chore: vendor selva_core from jnwnlee/selva@d7d40a9

Pure PyTorch SelVA source for SelvaModelLoader/FeatureExtractor/Sampler nodes.
Imports rewritten from selva.* to selva_core.*. mel_converter.py: replaced
librosa.filters.mel with pure-numpy implementation to avoid librosa→numba→NumPy
version incompatibility in some ComfyUI environments.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

selva_core/__init__.py

@@ -0,0 +1,3 @@
+# Vendored from https://github.com/jnwnlee/selva
+# Pinned commit: d7d40a992aab58e7cf246055681a657e5d8b4a4d
+# Imports rewritten from selva.* → selva_core.*

selva_core/data/av_utils.py

@@ -0,0 +1,190 @@
+import logging
+from dataclasses import dataclass
+from fractions import Fraction
+from pathlib import Path
+from typing import Optional
+
+import av
+import numpy as np
+import torch
+from av import AudioFrame
+
+log = logging.getLogger()
+
+@dataclass
+class VideoInfo:
+    duration_sec: float
+    fps: Fraction
+    clip_frames: torch.Tensor
+    sync_frames: torch.Tensor
+    all_frames: Optional[list[np.ndarray]]
+
+    @property
+    def height(self):
+        return self.all_frames[0].shape[0]
+
+    @property
+    def width(self):
+        return self.all_frames[0].shape[1]
+
+    @classmethod
+    def from_image_info(cls, image_info: 'ImageInfo', duration_sec: float,
+                        fps: Fraction) -> 'VideoInfo':
+        num_frames = int(duration_sec * fps)
+        all_frames = [image_info.original_frame] * num_frames
+        return cls(duration_sec=duration_sec,
+                   fps=fps,
+                   clip_frames=image_info.clip_frames,
+                   sync_frames=image_info.sync_frames,
+                   all_frames=all_frames)
+
+
+@dataclass
+class ImageInfo:
+    clip_frames: torch.Tensor
+    sync_frames: torch.Tensor
+    original_frame: Optional[np.ndarray]
+
+    @property
+    def height(self):
+        return self.original_frame.shape[0]
+
+    @property
+    def width(self):
+        return self.original_frame.shape[1]
+
+
+def read_frames(video_path: Path, list_of_fps: list[float], start_sec: float, end_sec: float,
+                need_all_frames: bool) -> tuple[list[np.ndarray], list[np.ndarray], Fraction]:
+    output_frames = [[] for _ in list_of_fps]
+    next_frame_time_for_each_fps = [0.0 for _ in list_of_fps]
+    time_delta_for_each_fps = [1 / fps for fps in list_of_fps]
+    all_frames = []
+
+    # container = av.open(video_path)
+    with av.open(video_path) as container:
+        stream = container.streams.video[0]
+        fps = stream.guessed_rate
+        stream.thread_type = 'AUTO'
+        for packet in container.demux(stream):
+            for frame in packet.decode():
+                frame_time = frame.time
+                if frame_time < start_sec:
+                    continue
+                if frame_time > end_sec:
+                    break
+
+                frame_np = None
+                if need_all_frames:
+                    frame_np = frame.to_ndarray(format='rgb24')
+                    all_frames.append(frame_np)
+
+                for i, _ in enumerate(list_of_fps):
+                    this_time = frame_time
+                    while this_time >= next_frame_time_for_each_fps[i]:
+                        if frame_np is None:
+                            frame_np = frame.to_ndarray(format='rgb24')
+
+                        output_frames[i].append(frame_np)
+                        next_frame_time_for_each_fps[i] += time_delta_for_each_fps[i]
+
+    output_frames = [np.stack(frames) for frames in output_frames]
+    return output_frames, all_frames, fps
+
+
+def normalize_video_chunk(video_chunk: torch.Tensor,
+                          expected_length: int,
+                          *,
+                          n_tolerance_frame: int = 1,
+                          desc: str = "") \
+    -> torch.Tensor:
+    # video_chunk: [T, H, W, C]
+    if video_chunk.shape[0] < expected_length:
+        if expected_length - video_chunk.shape[0] <= n_tolerance_frame:
+            # copy the last frame to make it the right length
+            log.warning(f'Video too short {desc}, padding {expected_length - video_chunk.shape[0]} frames with the last frame')
+            video_chunk = torch.cat([video_chunk, video_chunk[-1:].repeat(expected_length - video_chunk.shape[0], 1, 1, 1)])
+            assert video_chunk.shape[0] == expected_length
+        else:
+            raise RuntimeError(
+                f'Video too short {desc}, expected {expected_length}, got {video_chunk.shape[0]}'
+            )
+    video_chunk = video_chunk[:expected_length]
+    if video_chunk.shape[0] != expected_length:
+        raise RuntimeError(f'Video wrong length {desc}, '
+                            f'expected {expected_length}, '
+                            f'got {video_chunk.shape[0]}')
+    
+    return video_chunk
+
+
+def reencode_with_audio(video_info: VideoInfo, output_path: Path, audio: torch.Tensor,
+                        sampling_rate: int):
+    container = av.open(output_path, 'w')
+    output_video_stream = container.add_stream('h264', video_info.fps)
+    output_video_stream.codec_context.bit_rate = 10 * 1e6  # 10 Mbps
+    output_video_stream.width = video_info.width
+    output_video_stream.height = video_info.height
+    output_video_stream.pix_fmt = 'yuv420p'
+
+    output_audio_stream = container.add_stream('aac', sampling_rate)
+
+    # encode video
+    for image in video_info.all_frames:
+        image = av.VideoFrame.from_ndarray(image)
+        packet = output_video_stream.encode(image)
+        container.mux(packet)
+
+    for packet in output_video_stream.encode():
+        container.mux(packet)
+
+    # convert float tensor audio to numpy array
+    audio_np = audio.numpy().astype(np.float32)
+    audio_frame = AudioFrame.from_ndarray(audio_np, format='flt', layout='mono')
+    audio_frame.sample_rate = sampling_rate
+
+    for packet in output_audio_stream.encode(audio_frame):
+        container.mux(packet)
+
+    for packet in output_audio_stream.encode():
+        container.mux(packet)
+
+    container.close()
+
+
+def remux_with_audio(video_path: Path, audio: torch.Tensor, output_path: Path, sampling_rate: int):
+    """
+    NOTE: I don't think we can get the exact video duration right without re-encoding
+    so we are not using this but keeping it here for reference
+    """
+    video = av.open(video_path)
+    output = av.open(output_path, 'w')
+    input_video_stream = video.streams.video[0]
+    output_video_stream = output.add_stream(template=input_video_stream)
+    output_audio_stream = output.add_stream('aac', sampling_rate)
+
+    duration_sec = audio.shape[-1] / sampling_rate
+
+    for packet in video.demux(input_video_stream):
+        # We need to skip the "flushing" packets that `demux` generates.
+        if packet.dts is None:
+            continue
+        # We need to assign the packet to the new stream.
+        packet.stream = output_video_stream
+        output.mux(packet)
+
+    # convert float tensor audio to numpy array
+    audio_np = audio.numpy().astype(np.float32)
+    audio_frame = av.AudioFrame.from_ndarray(audio_np, format='flt', layout='mono')
+    audio_frame.sample_rate = sampling_rate
+
+    for packet in output_audio_stream.encode(audio_frame):
+        output.mux(packet)
+
+    for packet in output_audio_stream.encode():
+        output.mux(packet)
+
+    video.close()
+    output.close()
+
+    output.close()

selva_core/data/data_setup.py

@@ -0,0 +1,227 @@
+import logging
+import random
+from typing import Optional
+
+import numpy as np
+import torch
+from omegaconf import DictConfig, open_dict
+from torch.utils.data import DataLoader, Dataset
+from torch.utils.data.dataloader import default_collate
+from torch.utils.data.distributed import DistributedSampler
+
+from selva_core.data.vgg_sound import VGGSound
+from selva_core.data.eval.eval_video_dataset import VGGSound as VGGSoundEval
+from selva_core.data.eval.eval_video_dataset import InferenceVideoData, VGGMonoAudioBench
+from selva_core.data.eval.audiocaps import AudioCapsData
+from selva_core.data.mm_dataset import MultiModalDataset
+from selva_core.data.mixup import DataMixupCollate
+from selva_core.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+
+# Re-seed randomness every time we start a worker
+def worker_init_fn(worker_id: int):
+    worker_seed = torch.initial_seed() % (2**31) + worker_id + local_rank * 1000
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
+    log.debug(f'Worker {worker_id} re-seeded with seed {worker_seed} in rank {local_rank}')
+
+
+def load_video_data(cfg: DictConfig, data_cfg: DictConfig, normalize_audio: bool = False,
+                    ) -> Dataset:
+    dataset = VGGSound(root=data_cfg.root,
+                    tsv_path=data_cfg.subset_name,
+                    sample_rate=16_000,
+                    duration_sec=8.0,
+                    normalize_audio=normalize_audio,
+                    mmap_dir=data_cfg.memmap_dir,
+                    tsv_tsynch_path=data_cfg.tsv_tsynch,
+                    mmap_tsync_dir=data_cfg.memmap_dir_tsynch,
+                    data_dim=cfg.data_dim
+                    )
+
+    return dataset
+
+
+def load_audio_data(cfg: DictConfig, data_cfg: DictConfig) -> Dataset:
+    raise NotImplementedError('Audio data loading is not implemented yet')
+
+
+def setup_training_datasets(cfg: DictConfig, 
+                            generator: torch.Generator,
+                            ) -> tuple[Dataset, DistributedSampler, DataLoader]:
+    if cfg.mini_train:
+        vgg = load_video_data(cfg, cfg.data.VGGSound_val, normalize_audio=True)
+        dataset = MultiModalDataset([vgg], [])
+    if cfg.example_train:
+        video = load_video_data(cfg, cfg.data.Example_video, normalize_audio=True)
+        dataset = MultiModalDataset([video], [])
+    else:
+        vgg = load_video_data(cfg, cfg.data.VGGSound, normalize_audio=True)
+        # load the largest one first
+        # you can add more video/audio data upon demand, such as
+        # clotho = load_audio_data(cfg, cfg.data.Clotho)
+        dataset = MultiModalDataset([vgg], [])
+
+    batch_size = cfg.batch_size
+    num_workers = cfg.num_workers
+    pin_memory = cfg.pin_memory
+
+    if cfg.mixup.domain == 'data':
+        mixup_params = cfg.mixup.params
+        collate_fn = DataMixupCollate(generator=generator, 
+                                      **mixup_params)
+    else:
+        collate_fn = None
+
+    sampler, loader = construct_loader(dataset,
+                                       batch_size,
+                                       num_workers,
+                                       shuffle=True,
+                                       drop_last=True,
+                                       pin_memory=pin_memory,
+                                       collate_fn=collate_fn)
+
+    return dataset, sampler, loader
+
+
+def setup_test_datasets(cfg: DictConfig,
+                        generator: torch.Generator,
+                        ) -> tuple[Dataset, DistributedSampler, DataLoader]:
+    if cfg.example_train:
+        dataset = load_video_data(cfg, cfg.data.Example_video, normalize_audio=False, split='test')
+    elif cfg.dataset.startswith('vggsound'):
+        dataset = load_video_data(cfg, cfg.data.VGGSound_test, normalize_audio=False, split='test')
+    else:
+        raise NotImplementedError(f'Unknown dataset for test: {cfg.dataset}')
+
+    batch_size = cfg.batch_size
+    num_workers = cfg.get('num_workers_val', cfg.num_workers)
+    pin_memory = cfg.pin_memory
+    
+    if cfg.mixup.domain == 'data':
+        mixup_config = cfg.mixup.params
+        collate_fn = DataMixupCollate(generator=generator, 
+                                      **mixup_config)
+    else:
+        collate_fn = None
+    
+    sampler, loader = construct_loader(dataset,
+                                       batch_size,
+                                       num_workers,
+                                       shuffle=False,
+                                       drop_last=False,
+                                       pin_memory=pin_memory,
+                                       collate_fn=collate_fn)
+
+    return dataset, sampler, loader
+
+
+def setup_val_datasets(cfg: DictConfig,
+                       generator: torch.Generator,
+                       ) -> tuple[Dataset, DataLoader, DataLoader]:
+    if cfg.example_train:
+        dataset = load_video_data(cfg, cfg.data.Example_video, normalize_audio=False)
+    else:
+        dataset = load_video_data(cfg, cfg.data.VGGSound_val, normalize_audio=False)
+
+    val_batch_size = cfg.batch_size
+    val_eval_batch_size = cfg.eval_batch_size
+    num_workers = cfg.get('num_workers_val', cfg.num_workers)
+    pin_memory = cfg.pin_memory
+    
+    if cfg.mixup.domain == 'data':
+        mixup_config = cfg.mixup.params
+        collate_fn = DataMixupCollate(generator=generator, 
+                                      **mixup_config)
+    else:
+        collate_fn = None
+    
+    _, val_loader = construct_loader(dataset,
+                                     val_batch_size,
+                                     num_workers,
+                                     shuffle=False,
+                                     drop_last=False,
+                                     pin_memory=pin_memory,
+                                     collate_fn=collate_fn)
+    _, eval_loader = construct_loader(dataset,
+                                      val_eval_batch_size,
+                                      num_workers,
+                                      shuffle=False,
+                                      drop_last=False,
+                                      pin_memory=pin_memory,
+                                      collate_fn=collate_fn)
+
+    return dataset, val_loader, eval_loader
+
+
+def setup_eval_dataset(dataset_name: str, cfg: DictConfig) -> tuple[Dataset, DataLoader]:
+    if dataset_name.startswith('audiocaps_full'):
+        dataset = AudioCapsData(cfg.eval_data.audiocaps_full.audio_path,
+                                cfg.eval_data.audiocaps_full.csv_path)
+    elif dataset_name.startswith('audiocaps'):
+        dataset = AudioCapsData(cfg.eval_data.audiocaps.audio_path,
+                                cfg.eval_data.audiocaps.csv_path)
+    elif dataset_name.startswith('vggsound'):
+        dataset = VGGSound(cfg.eval_data.vggsound.video_path,
+                           cfg.eval_data.vggsound.csv_path,
+                           duration_sec=cfg.duration_s)
+    elif dataset_name.startswith('infer_video'):
+        dataset = InferenceVideoData(cfg.eval_data.infer_video.video_path,
+                                     cfg.eval_data.infer_video.jsonl_path,
+                                     duration_sec=cfg.duration_s)
+        cfg.batch_size = 1
+    elif dataset_name.startswith('example_video'):
+        dataset = VGGSoundEval(cfg.eval_data.Example_video.video_path,
+                               cfg.eval_data.Example_video.csv_path,
+                               duration_sec=cfg.duration_s)
+    elif dataset_name in ['vgg_monoaudio_intra', 'vgg_monoaudio_inter']:
+        dataset = VGGMonoAudioBench(cfg.eval_data[dataset_name].video_path,
+                                    cfg.eval_data[dataset_name].csv_path,
+                                    duration_sec=cfg.duration_s)
+        
+    else:
+        raise ValueError(f'Invalid dataset name: {dataset_name}')
+
+    batch_size = cfg.batch_size
+    num_workers = cfg.num_workers
+    pin_memory = cfg.pin_memory
+    _, loader = construct_loader(dataset,
+                                 batch_size,
+                                 num_workers,
+                                 shuffle=False,
+                                 drop_last=False,
+                                 pin_memory=pin_memory,
+                                 error_avoidance=True)
+    return dataset, loader
+
+
+def error_avoidance_collate(batch):
+    # Filter our None values
+    batch = [item for item in batch if item is not None]
+    if len(batch) == 0:
+        return None
+    return default_collate(batch)
+
+
+def construct_loader(dataset: Dataset,
+                     batch_size: int,
+                     num_workers: int,
+                     *,
+                     shuffle: bool = True,
+                     drop_last: bool = True,
+                     pin_memory: bool = False,
+                     error_avoidance: bool = False,
+                     collate_fn = None) -> tuple[DistributedSampler, DataLoader]:
+    train_sampler = DistributedSampler(dataset, rank=local_rank, shuffle=shuffle)
+    train_loader = DataLoader(dataset,
+                              batch_size,
+                              sampler=train_sampler,
+                              num_workers=num_workers,
+                              worker_init_fn=worker_init_fn,
+                              drop_last=drop_last,
+                              persistent_workers=num_workers > 0,
+                              pin_memory=pin_memory,
+                              collate_fn=error_avoidance_collate if error_avoidance else collate_fn)
+    return train_sampler, train_loader

selva_core/data/eval/audiocaps.py

@@ -0,0 +1,39 @@
+import logging
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Union
+
+import pandas as pd
+import torch
+from torch.utils.data.dataset import Dataset
+
+log = logging.getLogger()
+
+
+class AudioCapsData(Dataset):
+
+    def __init__(self, audio_path: Union[str, Path], csv_path: Union[str, Path]):
+        df = pd.read_csv(csv_path).to_dict(orient='records')
+
+        audio_files = sorted(os.listdir(audio_path))
+        audio_files = set(
+            [Path(f).stem for f in audio_files if f.endswith('.wav') or f.endswith('.flac')])
+
+        self.data = []
+        for row in df:
+            self.data.append({
+                'name': row['name'],
+                'caption': row['caption'],
+            })
+
+        self.audio_path = Path(audio_path)
+        self.csv_path = Path(csv_path)
+
+        log.info(f'Found {len(self.data)} matching audio files in {self.audio_path}')
+
+    def __getitem__(self, idx: int) -> torch.Tensor:
+        return self.data[idx]
+
+    def __len__(self):
+        return len(self.data)

selva_core/data/eval/eval_video_dataset.py

@@ -0,0 +1,237 @@
+import json
+import logging
+import os
+from pathlib import Path
+from typing import Union
+
+import pandas as pd
+import torch
+from torch.utils.data.dataset import Dataset
+from torchvision.transforms import v2
+from torio.io import StreamingMediaDecoder
+
+from selva_core.data.av_utils import normalize_video_chunk
+from selva_core.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+class VideoDataset(Dataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        *,
+        duration_sec: float = 8.0,
+        clip_video_required: bool = False,
+    ):
+        self.video_root = Path(video_root)
+        self.duration_sec = duration_sec
+        self.clip_video_required = clip_video_required
+
+        self.sync_expected_length = int(_SYNC_FPS * self.duration_sec)
+        self.sync_transform = v2.Compose([
+            v2.Resize((_SYNC_SIZE, _SYNC_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+            # v2.CenterCrop(_SYNC_SIZE),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+        
+        if self.clip_video_required:
+            self.clip_expected_length = int(_CLIP_FPS * self.duration_sec)
+            self.clip_transform = v2.Compose([
+                v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+                v2.ToImage(),
+                v2.ToDtype(torch.float32, scale=True),
+            ])
+
+        # to be implemented by subclasses
+        self.captions = {}
+        self.negative_captions = {}
+        self.videos = sorted(list(self.captions.keys()))
+
+    def sample(self, idx: int) -> dict[str, torch.Tensor]:
+        video_id = self.videos[idx]
+        caption = self.captions[video_id]
+        negative_caption = self.negative_captions.get(video_id, None)
+
+        reader = StreamingMediaDecoder(self.video_root / (video_id + '.mp4'))
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_SYNC_FPS * self.duration_sec),
+            frame_rate=_SYNC_FPS,
+            format='rgb24',
+        )
+        if self.clip_video_required:
+            reader.add_basic_video_stream(
+                frames_per_chunk=int(_CLIP_FPS * self.duration_sec),
+                frame_rate=_CLIP_FPS,
+                format='rgb24',
+            )
+
+        reader.fill_buffer()
+        data_chunk = reader.pop_chunks()
+
+        sync_chunk = data_chunk[0]
+        if sync_chunk is None:
+            raise RuntimeError(f'Sync video returned None {video_id}')
+        sync_chunk = normalize_video_chunk(sync_chunk, self.sync_expected_length, 
+                                           n_tolerance_frame=3, desc=video_id)
+        sync_chunk = self.sync_transform(sync_chunk)
+        
+        if self.clip_video_required:
+            clip_chunk = data_chunk[1]
+            if clip_chunk is None:
+                raise RuntimeError(f'CLIP video returned None {video_id}')
+            clip_chunk = normalize_video_chunk(clip_chunk, self.clip_expected_length, 
+                                               n_tolerance_frame=1, desc=video_id)
+            clip_chunk = self.clip_transform(clip_chunk)
+
+        data = {
+            'name': video_id,
+            'caption': caption,
+            'sync_video': sync_chunk,
+        }
+        if self.clip_video_required:
+            data['clip_video'] = clip_chunk
+        if negative_caption is not None:
+            data['negative_caption'] = negative_caption
+
+        return data
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        try:
+            return self.sample(idx)
+        except Exception as e:
+            log.error(f'Error loading video {self.videos[idx]}: {e}')
+            return None
+
+    def __len__(self):
+        return len(self.captions)
+
+
+class VGGSound(VideoDataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        csv_path: Union[str, Path],
+        *,
+        duration_sec: float = 8.0,
+        clip_video_required: bool = False,
+    ):
+        super().__init__(video_root, duration_sec=duration_sec, 
+                         clip_video_required=clip_video_required)
+        self.video_root = Path(video_root)
+        self.csv_path = Path(csv_path)
+
+        videos = sorted(os.listdir(self.video_root))
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+        self.captions = {}
+
+        df = pd.read_csv(csv_path, header=None, names=['id', 'sec', 'caption',
+                                                       'split']).to_dict(orient='records')
+
+        videos_no_found = []
+        for row in df:
+            if row['split'] == 'test':
+                start_sec = int(row['sec'])
+                video_id = str(row['id'])
+                # this is how our videos are named
+                video_name = f'{video_id}_{start_sec:06d}'
+                if video_name + '.mp4' not in videos:
+                    videos_no_found.append(video_name)
+                    continue
+
+                self.captions[video_name] = row['caption']
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+            log.info(f'{len(self.captions)} useable videos found')
+            if videos_no_found:
+                log.info(f'{len(videos_no_found)} found in {csv_path} but not in {video_root}')
+                log.info(
+                    'A small amount is expected, as not all videos are still available on YouTube')
+
+        self.videos = sorted(list(self.captions.keys()))
+
+
+class InferenceVideoData(VideoDataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        jsonl_root: Union[str, Path],
+        *,
+        duration_sec: float = 10.0,
+        clip_video_required: bool = False,
+    ):
+        super().__init__(video_root, duration_sec=duration_sec, 
+                         clip_video_required=clip_video_required)
+        self.video_root = Path(video_root)
+        self.jsonl_root = Path(jsonl_root)
+
+        videos = sorted(os.listdir(self.video_root))
+        videos = [v[:-4] for v in videos]  # remove extensions
+        self.captions = {}
+
+        for v in videos:
+            with open(self.jsonl_root / (v + '.jsonl')) as f:
+                data = json.load(f)
+                self.captions[v] = data['audio_prompt']
+                self.negative_captions[v] = data.get('negative_audio_prompt', None)
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+
+        self.videos = videos
+
+
+class VGGMonoAudioBench(VideoDataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        csv_path: Union[str, Path],
+        *,
+        duration_sec: float = 8.0,
+        clip_video_required: bool = False,
+    ):
+        super().__init__(video_root, duration_sec=duration_sec, 
+                         clip_video_required=clip_video_required)
+        self.video_root = Path(video_root)
+        self.csv_path = Path(csv_path)
+
+        videos = sorted(os.listdir(self.video_root))
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+        self.captions = {}
+        self.negative_captions = {}
+
+        df = pd.read_csv(csv_path, header=0, usecols=['file_name', 'label', 'paired_label']
+                         ).to_dict(orient='records')
+
+        videos_no_found = []
+        for row in df:
+            video_name = str(Path(row['file_name']).stem)
+            if video_name + '.mp4' not in videos:
+                videos_no_found.append(video_name)
+                continue
+
+            self.captions[video_name] = row['label']
+            self.negative_captions[video_name] = row['paired_label']
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+            log.info(f'{len(self.captions)} useable videos found')
+            if videos_no_found:
+                log.info(f'{len(videos_no_found)} found in {csv_path} but not in {video_root}!')
+
+        self.videos = sorted(list(self.captions.keys()))

selva_core/data/extraction/vgg_sound.py

@@ -0,0 +1,194 @@
+import logging
+import os
+from pathlib import Path
+from typing import Optional, Union
+
+import pandas as pd
+import torch
+import torchaudio
+from torch.utils.data.dataset import Dataset
+from torchvision.transforms import v2
+from torio.io import StreamingMediaDecoder
+
+from selva_core.data.av_utils import normalize_video_chunk
+from selva_core.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+class VGGSound(Dataset):
+
+    def __init__(
+        self,
+        root: Union[str, Path],
+        *,
+        tsv_path: Union[str, Path] = 'sets/vgg3-train.tsv',
+        audio_required: bool = True,
+        sample_rate: int = 16_000,
+        duration_sec: float = 8.0,
+        audio_samples: Optional[int] = None,
+        normalize_audio: bool = False,
+        clip_video_required: bool = True,
+    ):
+        self.root = Path(root)
+        self.audio_required = audio_required
+        if audio_required:
+            self.normalize_audio = normalize_audio
+            if audio_samples is None:
+                self.audio_samples = int(sample_rate * duration_sec)
+            else:
+                self.audio_samples = audio_samples
+                effective_duration = audio_samples / sample_rate
+                # make sure the duration is close enough, within 15ms
+                assert abs(effective_duration - duration_sec) < 0.015, \
+                    f'audio_samples {audio_samples} does not match duration_sec {duration_sec}'
+        self.clip_video_required = clip_video_required
+        
+        videos = sorted(os.listdir(self.root))
+        videos = set([Path(v).stem for v in videos])  # remove extensions
+        self.labels = {}
+        self.videos = []
+        missing_videos = []
+
+        # read the tsv for subset information
+        df_list = pd.read_csv(tsv_path, sep='\t', dtype={'id': str}).to_dict('records')
+        for record in df_list:
+            id = record['id']
+            label = record['label']
+            if id in videos:
+                self.labels[id] = label
+                self.videos.append(id)
+            else:
+                missing_videos.append(id)
+        
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {root}')
+            log.info(f'{len(self.videos)} videos found in {tsv_path}')
+            log.info(f'{len(missing_videos)} videos missing in {root}')
+
+        self.sample_rate = sample_rate
+        self.duration_sec = duration_sec
+
+        if audio_required:
+            self.expected_audio_length = self.audio_samples
+        self.sync_expected_length = int(_SYNC_FPS * self.duration_sec)
+        if clip_video_required:
+            self.clip_expected_length = int(_CLIP_FPS * self.duration_sec)
+
+        self.sync_transform = v2.Compose([
+            v2.Resize((_SYNC_SIZE, _SYNC_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+            # v2.CenterCrop(_SYNC_SIZE),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+        
+        if clip_video_required:
+            self.clip_transform = v2.Compose([
+                v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+                v2.ToImage(),
+                v2.ToDtype(torch.float32, scale=True),
+            ])    
+        if audio_required:
+            self.resampler = {}
+
+    def sample(self, idx: int) -> dict[str, torch.Tensor]:
+        video_id = self.videos[idx]
+        
+        label = self.labels[video_id]
+            
+        reader = StreamingMediaDecoder(self.root / (video_id + '.mp4'))
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_SYNC_FPS * self.duration_sec),
+            frame_rate=_SYNC_FPS,
+            format='rgb24',
+        )
+        if self.audio_required:
+            reader.add_basic_audio_stream(frames_per_chunk=2**30, )
+        if self.clip_video_required:
+            reader.add_basic_video_stream(
+                frames_per_chunk=int(_CLIP_FPS * self.duration_sec),
+                frame_rate=_CLIP_FPS,
+                format='rgb24',
+            )
+
+        reader.fill_buffer()
+        data_chunk = reader.pop_chunks()
+        
+        sync_chunk = data_chunk[0]
+        if sync_chunk is None:
+            raise RuntimeError(f'Sync video returned None {video_id}')
+        sync_chunk = normalize_video_chunk(sync_chunk, self.sync_expected_length, 
+                                           n_tolerance_frame=3, desc=video_id)
+        sync_chunk = self.sync_transform(sync_chunk)
+        
+        if self.audio_required:
+            audio_chunk = data_chunk[1]
+        
+        if self.clip_video_required:
+            clip_chunk = data_chunk[2 if self.audio_required else 1]
+            if clip_chunk is None:
+                raise RuntimeError(f'CLIP video returned None {video_id}')
+            clip_chunk = normalize_video_chunk(clip_chunk, self.clip_expected_length,
+                                               n_tolerance_frame=1, desc=video_id)
+            clip_chunk = self.clip_transform(clip_chunk)
+
+        # process audio
+        if self.audio_required:
+            sample_rate = int(reader.get_out_stream_info(1).sample_rate)
+            audio_chunk = audio_chunk.transpose(0, 1)
+            audio_chunk = audio_chunk.mean(dim=0)  # mono
+            if self.normalize_audio:
+                abs_max = audio_chunk.abs().max()
+                audio_chunk = audio_chunk * (0.95 / abs_max)
+                if abs_max <= 1e-6:
+                    raise RuntimeError(f'Audio is silent {video_id}')
+
+            # resample
+            if sample_rate == self.sample_rate:
+                audio_chunk = audio_chunk
+            else:
+                if sample_rate not in self.resampler:
+                    # https://pytorch.org/audio/stable/tutorials/audio_resampling_tutorial.html#kaiser-best
+                    self.resampler[sample_rate] = torchaudio.transforms.Resample(
+                        sample_rate,
+                        self.sample_rate,
+                        lowpass_filter_width=64,
+                        rolloff=0.9475937167399596,
+                        resampling_method='sinc_interp_kaiser',
+                        beta=14.769656459379492,
+                    )
+                audio_chunk = self.resampler[sample_rate](audio_chunk)
+
+            if audio_chunk.shape[0] < self.expected_audio_length:
+                raise RuntimeError(f'Audio too short {video_id}')
+            audio_chunk = audio_chunk[:self.expected_audio_length]
+
+        data = {
+            'id': video_id,
+            'caption': label,
+            'sync_video': sync_chunk,
+        }
+        
+        if self.audio_required:
+            data['audio'] = audio_chunk
+        if self.clip_video_required:
+            data['clip_video'] = clip_chunk
+
+        return data
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        try:
+            return self.sample(idx)
+        except Exception as e:
+            log.error(f'Error loading video {self.videos[idx]}: {e}')
+            return None
+
+    def __len__(self):
+        return len(self.labels)

selva_core/data/extraction/wav_dataset.py

@@ -0,0 +1,129 @@
+import logging
+import os
+from pathlib import Path
+from typing import Union
+
+import open_clip
+import pandas as pd
+import torch
+import torchaudio
+from torch.utils.data.dataset import Dataset
+
+log = logging.getLogger()
+
+
+class WavTextClipsDataset(Dataset):
+
+    def __init__(
+        self,
+        root: Union[str, Path],
+        *,
+        captions_tsv: Union[str, Path],
+        clips_tsv: Union[str, Path],
+        sample_rate: int,
+        num_samples: int,
+        normalize_audio: bool = False,
+        reject_silent: bool = False,
+        tokenizer_id: str = 'ViT-H-14-378-quickgelu',
+    ):
+        self.root = Path(root)
+        self.sample_rate = sample_rate
+        self.num_samples = num_samples
+        self.normalize_audio = normalize_audio
+        self.reject_silent = reject_silent
+        self.tokenizer = open_clip.get_tokenizer(tokenizer_id)
+
+        audios = sorted(os.listdir(self.root))
+        audios = set([
+            Path(audio).stem for audio in audios
+            if audio.endswith('.wav') or audio.endswith('.flac')
+        ])
+        self.captions = {}
+
+        # read the caption tsv
+        df_list = pd.read_csv(captions_tsv, sep='\t', dtype={'id': str}).to_dict('records')
+        for record in df_list:
+            id = record['id']
+            caption = record['caption']
+            self.captions[id] = caption
+
+        # read the clip tsv
+        df_list = pd.read_csv(clips_tsv, sep='\t', dtype={
+            'id': str,
+            'name': str
+        }).to_dict('records')
+        self.clips = []
+        for record in df_list:
+            record['id'] = record['id']
+            record['name'] = record['name']
+            id = record['id']
+            name = record['name']
+            record['caption'] = self.captions[name]
+            self.clips.append(record)
+
+        log.info(f'Found {len(self.clips)} audio files in {self.root}')
+
+        self.resampler = {}
+
+    def __getitem__(self, idx: int) -> torch.Tensor:
+        try:
+            clip = self.clips[idx]
+            audio_name = clip['name']
+            audio_id = clip['id']
+            caption = clip['caption']
+            start_sample = clip['start_sample']
+            end_sample = clip['end_sample']
+
+            audio_path = self.root / f'{audio_name}.flac'
+            if not audio_path.exists():
+                audio_path = self.root / f'{audio_name}.wav'
+                assert audio_path.exists()
+
+            audio_chunk, sample_rate = torchaudio.load(audio_path)
+            audio_chunk = audio_chunk.mean(dim=0)  # mono
+            abs_max = audio_chunk.abs().max()
+            if self.normalize_audio:
+                audio_chunk = audio_chunk / abs_max * 0.95
+
+            if self.reject_silent and abs_max < 1e-6:
+                log.warning(f'Rejecting silent audio')
+                return None
+
+            audio_chunk = audio_chunk[start_sample:end_sample]
+
+            # resample
+            if sample_rate == self.sample_rate:
+                audio_chunk = audio_chunk
+            else:
+                if sample_rate not in self.resampler:
+                    # https://pytorch.org/audio/stable/tutorials/audio_resampling_tutorial.html#kaiser-best
+                    self.resampler[sample_rate] = torchaudio.transforms.Resample(
+                        sample_rate,
+                        self.sample_rate,
+                        lowpass_filter_width=64,
+                        rolloff=0.9475937167399596,
+                        resampling_method='sinc_interp_kaiser',
+                        beta=14.769656459379492,
+                    )
+                audio_chunk = self.resampler[sample_rate](audio_chunk)
+
+            if audio_chunk.shape[0] < self.num_samples:
+                raise ValueError('Audio is too short')
+            audio_chunk = audio_chunk[:self.num_samples]
+
+            tokens = self.tokenizer([caption])[0]
+
+            output = {
+                'waveform': audio_chunk,
+                'id': audio_id,
+                'caption': caption,
+                'tokens': tokens,
+            }
+
+            return output
+        except Exception as e:
+            log.error(f'Error reading {audio_path}: {e}')
+            return None
+
+    def __len__(self):
+        return len(self.clips)

selva_core/data/mixup.py

@@ -0,0 +1,338 @@
+""" Embedding Mixup
+Reference: https://github.com/huggingface/pytorch-image-models/blob/main/timm/data/mixup.py
+"""
+from typing import Literal, Tuple, Union, List, Optional
+from functools import partial
+import gc
+
+import numpy as np
+import torch
+from torch.utils.data.dataloader import default_collate
+from torchvision.transforms import v2
+from einops import rearrange
+from omegaconf import DictConfig
+
+from selva_core.data.vgg_sound import _SYNC_SIZE
+
+
+class MixupBase:
+    """ Base class for mixup on either data or feature domain.
+    Applies different params to each element or whole batch.
+
+    Args:
+        generator (Optional[torch.Generator]): Random number generator for reproducibility
+        modality (Literal['video', 'audio', 'both']): Modality to apply mixup on.
+        mixup_lambda (float): Mixup lambda value, mixup is active if in [0., 1.].
+        mixup_alpha (float): Mixup alpha value, mixup is active if > 0.
+        prob (float): Probability of applying mixup per batch or element
+        mode (Literal['elem','pair','batch', 'half']): How to apply mixup params (per 'batch', 'pair' (pair of elements), 'elem' (element), 'half' (half batch))
+        eps (float): Small epsilon value to avoid zero lambda
+    """
+    def __init__(self, generator:torch.Generator, 
+                 *, 
+                 modality:Literal['video', 'audio', 'both'],
+                 mixup_lambda:float=0.5, mixup_alpha:float=1., prob:float=1.0, 
+                 mode:Literal['elem','pair','batch', 'half']='batch',
+                 eps:float=0.05
+                 ):
+        self.modality = modality
+        self.mixup_lambda:float = mixup_lambda
+        self.mixup_alpha:float = mixup_alpha
+        self.mix_prob:float = prob
+        self.mode:str = mode
+        self.eps:float = eps
+        self.mixup_enabled:bool = True  # set to false to disable mixing (intended to be set by train loop)
+        if generator.device.type == 'cuda':
+            self.generator_cuda = generator
+            generator_seed = generator.initial_seed()
+            self.generator = torch.Generator(device='cpu')
+            self.generator.manual_seed(generator_seed)
+        else:
+            self.generator = generator
+
+        if not (self.mixup_lambda >= 0. and self.mixup_lambda <= 1.):
+            raise ValueError(f"mixup_lambda {self.mixup_lambda} should be in [0., 1.].")
+        if not self.mixup_alpha >= 0.:
+            raise ValueError(f"mixup_alpha {self.mixup_alpha} >= 0. should be true.")
+        if (self.mixup_alpha > 0. and self.mixup_lambda < 1.) or (self.mixup_alpha == 0. and self.mixup_lambda == 1.):
+            raise ValueError(f"One of mixup_alpha {self.mixup_alpha} > 0., mixup_lambda {self.mixup_lambda} < 1. should be true.")
+    
+    def _params_per_elem(self, batch_size:int) -> np.ndarray:
+        lam:np.ndarray = np.ones(batch_size, dtype=np.float32)
+        if self.mixup_enabled:
+            if self.mixup_lambda < 1.: # constant lambda
+                lam_mix = np.full(batch_size, self.mixup_lambda, dtype=np.float32)
+            elif self.mixup_alpha > 0.: # sampled lambda
+                # Use torch's beta distribution with generator
+                lam_mix = torch.distributions.Beta(
+                    torch.tensor([self.mixup_alpha]), 
+                    torch.tensor([self.mixup_alpha]),
+                ).sample([batch_size]).numpy().astype(np.float32).reshape(-1)
+            else:
+                assert False, f"One of mixup_alpha {self.mixup_alpha} > 0., mixup_lambda {self.mixup_lambda} < 1. should be true."
+            lam_mix[lam_mix < self.eps] = self.eps
+            
+            # Use torch's random with generator for the random comparison
+            rand_vals = torch.rand(batch_size, generator=self.generator).numpy()
+            lam = np.where(rand_vals < self.mix_prob, lam_mix, lam)
+        return lam
+
+    def _params_per_batch(self) -> float:
+        lam:float = 1.
+        if self.mixup_enabled:
+            if self.mixup_lambda < 1.: # constant lambda
+                lam = self.mixup_lambda
+            elif self.mixup_alpha > 0.: # sampled lambda
+                lam = torch.distributions.Beta(
+                    torch.tensor([self.mixup_alpha]), 
+                    torch.tensor([self.mixup_alpha]),
+                ).sample().item()
+            else:
+                assert False, f"mixup_alpha {self.mixup_alpha} > 0., mixup_lambda {self.mixup_lambda} < 1. should be true."
+            if lam < self.eps: lam = self.eps
+            lam = float(lam)
+        return lam
+
+
+class DataMixupCollate(MixupBase):
+    """ Mixup video in data domain.
+    Applies different params to each element or whole batch.
+
+    Args:
+        generator (Optional[torch.Generator]): Random number generator for reproducibility
+        modality (Literal['video', 'audio', 'both']): Modality to apply mixup on.
+        mixup_lambda (float): Mixup lambda value, mixup is active if in [0., 1.].
+        mixup_alpha (float): Mixup alpha value, mixup is active if > 0.
+        prob (float): Probability of applying mixup per batch or element
+        mode (Literal['elem','pair','batch', 'half']): How to apply mixup params (per 'batch', 'pair' (pair of elements), 'elem' (element), 'half' (half batch))
+        eps (float): Small epsilon value to avoid zero lambda
+    """
+    def __init__(self, generator:torch.Generator, 
+                 *, 
+                 modality:Literal['video', 'audio', 'both']='video',
+                 mixup_lambda:float=0.5, mixup_alpha:float=1., prob:float=1.0, 
+                 mode:Literal['elem','pair','batch', 'half']='batch',
+                 eps:float=0.05
+                 ):
+        super().__init__(generator, modality=modality,
+                         mixup_lambda=mixup_lambda, mixup_alpha=mixup_alpha, prob=prob, 
+                         mode=mode, eps=eps)
+        
+        self.source_video_key= 'sync_video'
+        self.source_audio_key = 'audio'
+        self.target_video_key = 'sync_video_mixed'
+        self.target_audio_key = 'audio_mixed'
+        
+        if not mode == 'batch':
+            raise ValueError(f"Mode {mode} is not supported for data domain.")
+        self.sync_transform = v2.Compose([
+            v2.CenterCrop(_SYNC_SIZE),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+    
+    def _concat_video_frames(self, batch:list, target_key:str='sync_video_mixed', source_key:str='sync_video') -> float:
+        # only batch mode supported
+        batch_size:int = len(batch)
+        lam:float = self._params_per_batch()
+        
+        if lam == 1.:
+            # no mixup, just return
+            for i in range(batch_size):
+                batch[i][target_key] = batch[i][source_key]
+            return lam
+        
+        # Randomly choose between horizontal and vertical resizing using
+        orig_size = int(lam * _SYNC_SIZE)
+        is_horizontal = True # torch.rand(1, generator=self.generator).item() < 0.5
+        if is_horizontal:
+            # Horizontal resize
+            resize_shape_orig = (_SYNC_SIZE, orig_size)
+            resize_shape_pair = (_SYNC_SIZE, _SYNC_SIZE-orig_size)    
+        else:
+            # Vertical resize
+            resize_shape_orig = (orig_size, _SYNC_SIZE)
+            resize_shape_pair = (_SYNC_SIZE-orig_size, _SYNC_SIZE)
+        sync_resize_orig = v2.Compose([
+            v2.Resize(resize_shape_orig, interpolation=v2.InterpolationMode.BICUBIC),
+        ])
+        sync_resize_pair = v2.Compose([
+            v2.Resize(resize_shape_pair, interpolation=v2.InterpolationMode.BICUBIC),
+        ])
+        
+        batch_videos_orig = torch.stack([batch[i][source_key] for i in range(batch_size)], dim=0)
+        batch_videos_pair = torch.stack([batch[batch_size - i - 1][source_key] for i in range(batch_size)], dim=0)
+        # (B, T, C, H, W)
+        # pass through resize, transform and concat
+        batch_videos_orig = sync_resize_orig(batch_videos_orig)
+        batch_videos_pair = sync_resize_pair(batch_videos_pair)
+        batch_videos_concat = torch.cat((batch_videos_orig, batch_videos_pair), dim=-1 if is_horizontal else -2)
+        batch_videos_concat = self.sync_transform(batch_videos_concat)
+
+        num_mixup = int(self.mix_prob * batch_size)
+        for i in range(num_mixup):
+            batch[i][target_key] = batch_videos_concat[i]
+        for i in range(num_mixup, batch_size):
+            batch[i][target_key] = batch[i][source_key] # no mixup
+            
+        del batch_videos_orig, batch_videos_pair, sync_resize_orig, sync_resize_pair
+        gc.collect()
+            
+        return lam 
+    
+    def _mix_audio_samples(self, batch:list, target_key:str='audio_mixed', source_key:str='audio',
+                           normalize:bool = True) -> float:
+        # assume source_key audios are normalized
+        batch_size:int = len(batch)
+        lam:float = self._params_per_batch()
+        
+        if lam == 1.:
+            # no mixup, just return
+            for i in range(batch_size):
+                batch[i][target_key] = batch[i][source_key]
+            return lam
+        
+        num_mixup = int(self.mix_prob * batch_size)
+        for i in range(num_mixup):
+            batch[i][target_key] = batch[i][source_key] * lam + batch[batch_size - i - 1][source_key] * (1 - lam)
+            if normalize:
+                source_abs_max = batch[i][source_key].abs().max()
+                target_abs_max = batch[i][target_key].abs().max()
+                batch[i][target_key] = batch[i][target_key] * (source_abs_max / target_abs_max)
+        for i in range(num_mixup, batch_size):
+            batch[i][target_key] = batch[i][source_key] # no mixup
+
+        return lam
+    
+    def __call__(self, batch:list, _=None) -> torch.tensor:
+        batch_size:int = len(batch)
+        assert batch_size % 2 == 0, f'Batch size {batch_size} should be even when using mixup'
+        half = 'half' in self.mode
+        if half:
+            batch_size //= 2
+        
+        if self.modality == 'video' or self.modality == 'both':
+            lam = self._concat_video_frames(batch, target_key=self.target_video_key, source_key=self.source_video_key)
+        if self.modality == 'audio' or self.modality == 'both':
+            # raise NotImplementedError('Audio mixup is not implemented yet.')
+            lam = self._mix_audio_samples(batch, target_key=self.target_audio_key, source_key=self.source_audio_key)
+        
+        return default_collate(batch)
+
+
+class FeatureMixup(MixupBase):
+    """ Mixup video in feature domain.
+    Applies different params to each element or whole batch.
+
+    Args:
+        generator (Optional[torch.Generator]): Random number generator for reproducibility
+        modality (Literal['video', 'audio', 'both']): Modality to apply mixup on.
+        mixup_lambda (float): Mixup lambda value, mixup is active if in [0., 1.].
+        mixup_alpha (float): Mixup alpha value, mixup is active if > 0.
+        prob (float): Probability of applying mixup per batch or element
+        mode (Literal['elem','pair','batch', 'half']): How to apply mixup params (per 'batch', 'pair' (pair of elements), 'elem' (element), 'half' (half batch))
+        eps (float): Small epsilon value to avoid zero lambda
+    """
+    def __init__(self, generator:torch.Generator, 
+                 *, 
+                 modality:Literal['video', 'audio', 'both']='video',
+                 mixup_lambda:float=0.5, mixup_alpha:float=1., prob:float=1.0, 
+                 mode:Literal['elem','pair','batch', 'half']='batch',
+                 eps:float=0.05
+                 ):
+        super().__init__(generator, modality=modality,
+                         mixup_lambda=mixup_lambda, mixup_alpha=mixup_alpha, prob=prob, 
+                         mode=mode, eps=eps)
+        self.source_video_key= 'sync_f_vid_orig'
+        self.source_audio_key = 'sync_f_aud_orig'
+        self.target_video_key = 'sync_f_vid_mixed'
+        self.target_audio_key = 'sync_f_aud_mixed'
+
+    def _mix_elem_collate(self, batch:dict, 
+                          target_keys:List[str]=['sync_features_mixed'], source_keys:List[str]=['sync_features_orig'],
+                          half:bool=False) -> torch.tensor:
+        assert len(target_keys) == len(source_keys), f"Length of target_keys {len(target_keys)} and source_keys {len(source_keys)} should be equal."
+        batch_size:int = len(batch['id'])
+        num_elem:int = batch_size // 2 if half else batch_size
+        lam_batch:torch.tensor = torch.from_numpy(self._params_per_elem(num_elem))
+        
+        indices = torch.arange(num_elem)
+        mix_indices = batch_size - indices - 1
+        mix_mask = lam_batch < 1
+        active_indices = indices[mix_mask]
+        active_mix_indices = mix_indices[mix_mask]
+        active_lambdas = lam_batch[mix_mask].unsqueeze(1)
+        for target_key, source_key in zip(target_keys, source_keys):
+            batch[target_key][active_indices] = (
+                batch[source_key][active_indices] * active_lambdas + 
+                batch[source_key][active_mix_indices] * (1 - active_lambdas)
+            )
+            batch[target_key][~indices[mix_mask]] = batch[source_key][~indices[mix_mask]]
+        if half:
+            lam_batch = torch.cat((lam_batch, torch.ones(num_elem, dtype=lam_batch.dtype)))
+        return lam_batch.unsqueeze(1)
+
+    def _mix_pair_collate(self, batch:dict,
+                          target_keys:List[str]=['sync_features_mixed'], source_keys:List[str]=['sync_features_orig']) -> torch.tensor:
+        assert len(target_keys) == len(source_keys), f"Length of target_keys {len(target_keys)} and source_keys {len(source_keys)} should be equal."
+        batch_size:int = len(batch['id'])
+        lam_batch:torch.tensor = torch.from_numpy(self._params_per_elem(batch_size // 2))
+        
+        indices = torch.arange(batch_size // 2)
+        mix_indices = batch_size - indices - 1
+        mix_mask = lam_batch < 1
+        active_indices = indices[mix_mask]
+        active_mix_indices = mix_indices[mix_mask]
+        active_lambdas = lam_batch[mix_mask].unsqueeze(1)
+        for target_key, source_key in zip(target_keys, source_keys):
+            batch[target_key][active_indices] = (
+                batch[source_key][active_indices] * active_lambdas + 
+                batch[source_key][active_mix_indices] * (1 - active_lambdas)
+            )
+            batch[target_key][active_mix_indices] = (
+                batch[source_key][active_mix_indices] * active_lambdas + 
+                batch[source_key][active_indices] * (1 - active_lambdas)
+            )
+            batch[target_key][~indices[mix_mask]] = batch[source_key][~indices[mix_mask]]
+            batch[target_key][~mix_indices[mix_mask]] = batch[source_key][~mix_indices[mix_mask]]
+        lam_batch = torch.cat((lam_batch, lam_batch.flip(0)))
+        return lam_batch.unsqueeze(1)
+
+    def _mix_batch_collate(self, batch:dict,
+                           target_keys:List[str]=['sync_features_mixed'], source_keys:List[str]=['sync_features_orig']) -> float:
+        assert len(target_keys) == len(source_keys), f"Length of target_keys {len(target_keys)} and source_keys {len(source_keys)} should be equal."
+        lam:float = self._params_per_batch()
+        
+        for target_key, source_key in zip(target_keys, source_keys):
+            num_mixup = int(self.mix_prob * batch[source_key].shape[0])
+            flipped_source = torch.flip(batch[source_key], dims=[0])
+            batch[target_key] = batch[source_key] * lam + flipped_source * (1 - lam)
+            batch[target_key][num_mixup:] = batch[source_key][num_mixup:] # no mixup
+        return lam
+    
+    def __call__(self, batch:dict, _=None) -> None:
+        batch_size:int = len(batch['id'])
+        assert batch_size % 2 == 0, f'Batch size(={batch_size}) should be even when using this'
+        half = 'half' in self.mode
+        if half:
+            batch_size //= 2
+        
+        # Mixup
+        if self.mode == 'elem' or self.mode == 'half':
+            collate_fn = partial(self._mix_elem_collate, half=half)
+        elif self.mode == 'pair':
+            collate_fn = self._mix_pair_collate
+        else:
+            collate_fn = self._mix_batch_collate
+        
+        if self.modality == 'both':
+            target_keys, source_keys = [self.target_video_key, self.target_audio_key], [self.source_video_key, self.source_audio_key]
+        elif self.modality == 'video':
+            target_keys, source_keys = [self.target_video_key], [self.source_video_key]
+        elif self.modality == 'audio':
+            target_keys, source_keys = [self.target_audio_key], [self.source_audio_key]
+        lam = collate_fn(batch, target_keys=target_keys, source_keys=source_keys)
+        
+        # return batch
+

selva_core/data/mm_dataset.py

@@ -0,0 +1,45 @@
+import bisect
+
+import torch
+from torch.utils.data.dataset import Dataset
+
+
+# modified from https://pytorch.org/docs/stable/_modules/torch/utils/data/dataset.html#ConcatDataset
+class MultiModalDataset(Dataset):
+    datasets: list[Dataset]
+    cumulative_sizes: list[int]
+
+    @staticmethod
+    def cumsum(sequence):
+        r, s = [], 0
+        for e in sequence:
+            l = len(e)
+            r.append(l + s)
+            s += l
+        return r
+
+    def __init__(self, video_datasets: list[Dataset], audio_datasets: list[Dataset]):
+        super().__init__()
+        self.video_datasets = list(video_datasets)
+        self.audio_datasets = list(audio_datasets)
+        self.datasets = self.video_datasets + self.audio_datasets
+
+        self.cumulative_sizes = self.cumsum(self.datasets)
+
+    def __len__(self):
+        return self.cumulative_sizes[-1]
+
+    def __getitem__(self, idx):
+        if idx < 0:
+            if -idx > len(self):
+                raise ValueError("absolute value of index should not exceed dataset length")
+            idx = len(self) + idx
+        dataset_idx = bisect.bisect_right(self.cumulative_sizes, idx)
+        if dataset_idx == 0:
+            sample_idx = idx
+        else:
+            sample_idx = idx - self.cumulative_sizes[dataset_idx - 1]
+        return self.datasets[dataset_idx][sample_idx]
+
+    def compute_latent_stats(self) -> tuple[torch.Tensor, torch.Tensor]:
+        return self.video_datasets[0].compute_latent_stats()

selva_core/data/utils.py

@@ -0,0 +1,148 @@
+import logging
+import os
+import random
+import tempfile
+from pathlib import Path
+from typing import Any, Optional, Union
+
+import torch
+import torch.distributed as dist
+from tensordict import MemoryMappedTensor
+from torch.utils.data import DataLoader
+from torch.utils.data.dataset import Dataset
+from tqdm import tqdm
+
+from selva_core.utils.dist_utils import local_rank, world_size
+
+scratch_path = Path(os.environ['SLURM_SCRATCH'] if 'SLURM_SCRATCH' in os.environ else '/dev/shm')
+shm_path = Path('/dev/shm')
+
+log = logging.getLogger()
+
+
+def reseed(seed):
+    random.seed(seed)
+    torch.manual_seed(seed)
+
+
+def local_scatter_torch(obj: Optional[Any]):
+    if world_size == 1:
+        # Just one worker. Do nothing.
+        return obj
+
+    array = [obj] * world_size
+    target_array = [None]
+    if local_rank == 0:
+        dist.scatter_object_list(target_array, scatter_object_input_list=array, src=0)
+    else:
+        dist.scatter_object_list(target_array, scatter_object_input_list=None, src=0)
+    return target_array[0]
+
+
+class ShardDataset(Dataset):
+
+    def __init__(self, root):
+        self.root = root
+        self.shards = sorted(os.listdir(root))
+
+    def __len__(self):
+        return len(self.shards)
+
+    def __getitem__(self, idx):
+        return torch.load(os.path.join(self.root, self.shards[idx]), weights_only=True)
+
+
+def get_tmp_dir(in_memory: bool) -> Path:
+    return shm_path if in_memory else scratch_path
+
+
+def load_shards_and_share(data_path: Union[str, Path], ids: list[int],
+                          in_memory: bool) -> MemoryMappedTensor:
+    if local_rank == 0:
+        with tempfile.NamedTemporaryFile(prefix='shared-tensor-', dir=get_tmp_dir(in_memory)) as f:
+            log.info(f'Loading shards from {data_path} into {f.name}...')
+            data = load_shards(data_path, ids=ids, tmp_file_path=f.name)
+            data = share_tensor_to_all(data)
+            torch.distributed.barrier()
+            f.close()  # why does the context manager not close the file for me?
+    else:
+        log.info('Waiting for the data to be shared with me...')
+        data = share_tensor_to_all(None)
+        torch.distributed.barrier()
+
+    return data
+
+
+def load_shards(
+    data_path: Union[str, Path],
+    ids: list[int],
+    *,
+    tmp_file_path: str,
+) -> Union[torch.Tensor, dict[str, torch.Tensor]]:
+
+    id_set = set(ids)
+    shards = sorted(os.listdir(data_path))
+    log.info(f'Found {len(shards)} shards in {data_path}.')
+    first_shard = torch.load(os.path.join(data_path, shards[0]), weights_only=True)
+
+    log.info(f'Rank {local_rank} created file {tmp_file_path}')
+    first_item = next(iter(first_shard.values()))
+    log.info(f'First item shape: {first_item.shape}')
+    mm_tensor = MemoryMappedTensor.empty(shape=(len(ids), *first_item.shape),
+                                         dtype=torch.float32,
+                                         filename=tmp_file_path,
+                                         existsok=True)
+    total_count = 0
+    used_index = set()
+    id_indexing = {i: idx for idx, i in enumerate(ids)}
+    # faster with no workers; otherwise we need to set_sharing_strategy('file_system')
+    loader = DataLoader(ShardDataset(data_path), batch_size=1, num_workers=0)
+    for data in tqdm(loader, desc='Loading shards'):
+        for i, v in data.items():
+            if i not in id_set:
+                continue
+
+            # tensor_index = ids.index(i)
+            tensor_index = id_indexing[i]
+            if tensor_index in used_index:
+                raise ValueError(f'Duplicate id {i} found in {data_path}.')
+            used_index.add(tensor_index)
+            mm_tensor[tensor_index] = v
+            total_count += 1
+
+    assert total_count == len(ids), f'Expected {len(ids)} tensors, got {total_count}.'
+    log.info(f'Loaded {total_count} tensors from {data_path}.')
+
+    return mm_tensor
+
+
+def share_tensor_to_all(x: Optional[MemoryMappedTensor]) -> MemoryMappedTensor:
+    """
+    x: the tensor to be shared; None if local_rank != 0
+    return: the shared tensor
+    """
+
+    # there is no need to share your stuff with anyone if you are alone; must be in memory
+    if world_size == 1:
+        return x
+
+    if local_rank == 0:
+        assert x is not None, 'x must not be None if local_rank == 0'
+    else:
+        assert x is None, 'x must be None if local_rank != 0'
+
+    if local_rank == 0:
+        filename = x.filename
+        meta_information = (filename, x.shape, x.dtype)
+    else:
+        meta_information = None
+
+    filename, data_shape, data_type = local_scatter_torch(meta_information)
+    if local_rank == 0:
+        data = x
+    else:
+        data = MemoryMappedTensor.from_filename(filename=filename,
+                                                dtype=data_type,
+                                                shape=data_shape)
+
+    return data

selva_core/data/vgg_sound.py

@@ -0,0 +1,299 @@
+import logging
+import os
+from pathlib import Path
+from typing import Optional, Union
+
+import pandas as pd
+import torch
+import torchaudio
+from torch.utils.data.dataset import Dataset
+from torchvision.transforms import v2
+from torio.io import StreamingMediaDecoder
+from tensordict import TensorDict
+
+from selva_core.data.av_utils import normalize_video_chunk
+from selva_core.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+class VGGSound(Dataset):
+
+    def __init__(
+        self,
+        root: Union[str, Path],
+        *,
+        tsv_path: Union[str, Path] = 'sets/vgg3-train.tsv',
+        for_generator: bool = True,
+        audio_required: bool = False,
+        sample_rate: int = 16_000,
+        duration_sec: float = 8.0,
+        audio_samples: Optional[int] = None,
+        normalize_audio: bool = False,
+        clip_video_required: bool = False,
+        mmap_dir: Union[str, Path] = None,
+        tsv_tsynch_path: Union[str, Path] = None,
+        mmap_tsync_dir: Union[str, Path] = None,
+        data_dim: dict[str, int] = None,
+    ):
+        self.root = Path(root)
+        self.audio_required = audio_required
+        if audio_required:
+            self.normalize_audio = normalize_audio
+            if audio_samples is None:
+                self.audio_samples = int(sample_rate * duration_sec)
+            else:
+                self.audio_samples = audio_samples
+                effective_duration = audio_samples / sample_rate
+                # make sure the duration is close enough, within 15ms
+                assert abs(effective_duration - duration_sec) < 0.015, \
+                    f'audio_samples {audio_samples} does not match duration_sec {duration_sec}'
+        self.clip_video_required = clip_video_required
+        self.for_generator = for_generator
+
+        videos = sorted(os.listdir(self.root))
+        videos = set([Path(v).stem for v in videos])  # remove extensions
+        self.labels = {}
+        self.videos = []
+        missing_videos = []
+
+        # read the tsv for subset information
+        df_list = pd.read_csv(tsv_path, sep='\t', dtype={'id': str}).to_dict('records')
+        for record in df_list:
+            id = record['id']
+            label = record['label']
+            if id in videos:
+                self.labels[id] = label
+                self.videos.append(id)
+            else:
+                missing_videos.append(id)
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {root}')
+            log.info(f'{len(self.videos)} videos found in {tsv_path}')
+            log.info(f'{len(missing_videos)} videos missing in {root}')
+
+        self.sample_rate = sample_rate
+        self.duration_sec = duration_sec
+
+        if audio_required:
+            self.expected_audio_length = self.audio_samples
+        self.sync_expected_length = int(_SYNC_FPS * self.duration_sec)
+        if clip_video_required:
+            self.clip_expected_length = int(_CLIP_FPS * self.duration_sec)
+
+        self.sync_transform = v2.Compose([
+            v2.Resize((_SYNC_SIZE, _SYNC_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+            # v2.CenterCrop(_SYNC_SIZE),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+        
+        if clip_video_required:
+            self.clip_transform = v2.Compose([
+                v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+                v2.ToImage(),
+                v2.ToDtype(torch.float32, scale=True),
+            ])    
+        if audio_required:
+            self.resampler = {}
+        
+        # mmap
+        log.info(f'Loading precomputed mmap from {mmap_dir}')
+        mmap_dir = Path(mmap_dir)
+        td = TensorDict.load_memmap(mmap_dir)
+        log.info(f'Loaded precomputed mmap from {mmap_dir}')
+        self.sync_features = td['sync_features']
+        if for_generator:
+            self.mean = td['mean']
+            self.std = td['std']
+            self.text_clip_features = td['text_features']
+        if clip_video_required:
+            self.clip_features = td['clip_features']
+        else:
+            self.clip_features = None
+        self.id2idx_mmap = {d['id']: i for i, d in enumerate(df_list)}
+        
+        mmap_tsync_dir = Path(mmap_tsync_dir)
+        td_tsync = TensorDict.load_memmap(mmap_tsync_dir)
+        log.info(f'Loaded precomputed tsync mmap from {mmap_tsync_dir}')
+        self.text_features = td_tsync['text_features']
+        self.text_masks = td_tsync['text_masks']
+        df_list_tsync = pd.read_csv(tsv_tsynch_path, sep='\t').to_dict('records')
+        self.id2idx_mmap_tsync = {d['id']: i for i, d in enumerate(df_list_tsync)}
+
+        if local_rank == 0:
+            log.info(f'Loaded {len(self)} samples.')
+            log.info(f'Loaded sync_features: {self.sync_features.shape}.')
+            log.info(f'Loaded text_features: {self.text_features.shape}.')
+            log.info(f'Loaded text_masks: {self.text_masks.shape}.')
+            if for_generator:
+                log.info(f'Loaded mean: {self.mean.shape}.')
+                log.info(f'Loaded std: {self.std.shape}.')
+                log.info(f'Loaded text_clip_features: {self.text_clip_features.shape}.')
+            if clip_video_required:
+                log.info(f'Loaded clip_features: {self.clip_features.shape}.')
+        
+        assert self.sync_features.shape[1] == data_dim['sync_seq_len'], \
+            f'{self.sync_features.shape[1]} != {data_dim["sync_seq_len"]}'
+        assert self.text_features.shape[1] <= data_dim['text_flant5_max_seq_len'], \
+            f'{self.text_features.shape[1]} > {data_dim["text_flant5_max_seq_len"]}'
+        assert self.text_masks.shape[1] <= data_dim['text_flant5_max_seq_len'], \
+            f'{self.text_masks.shape[1]} > {data_dim["text_flant5_max_seq_len"]}'
+        assert self.sync_features.shape[-1] == data_dim['sync_dim'], \
+            f'{self.sync_features.shape[-1]} != {data_dim["sync_dim"]}'
+        assert self.text_features.shape[-1] == data_dim['text_flant5_dim'], \
+            f'{self.text_features.shape[-1]} != {data_dim["text_flant5_dim"]}'
+        if for_generator:
+            assert self.mean.shape[1] == data_dim['latent_seq_len'], \
+                f'{self.mean.shape[1]} != {data_dim["latent_seq_len"]}'
+            assert self.std.shape[1] == data_dim['latent_seq_len'], \
+                f'{self.std.shape[1]} != {data_dim["latent_seq_len"]}'
+            assert self.text_clip_features.shape[1] == data_dim['text_clip_seq_len'], \
+                f'{self.text_clip_features.shape[1]} != {data_dim["text_clip_seq_len"]}'
+            assert self.text_clip_features.shape[-1] == data_dim['text_clip_dim'], \
+                f'{self.text_clip_features.shape[-1]} != {data_dim["text_clip_dim"]}'
+        if clip_video_required:
+            assert self.clip_features.shape[1] == data_dim['clip_seq_len'], \
+                f'{self.clip_features.shape[1]} != {data_dim["clip_seq_len"]}'
+            assert self.clip_features.shape[-1] == data_dim['clip_dim'], \
+                f'{self.clip_features.shape[-1]} != {data_dim["clip_dim"]}'
+            
+        self.video_exist = torch.tensor(1, dtype=torch.bool)
+        self.text_exist = torch.tensor(1, dtype=torch.bool)
+
+
+    def compute_latent_stats(self) -> tuple[torch.Tensor, torch.Tensor]: # mmap
+        latents = self.mean
+        return latents.mean(dim=(0, 1)), latents.std(dim=(0, 1))
+    
+    def get_memory_mapped_tensor(self) -> TensorDict:
+        td = TensorDict({
+            'sync_features': self.sync_features,
+            'text_features': self.text_features,
+            'text_masks': self.text_masks,
+        })
+        if self.for_generator:
+            td['mean'] = self.mean
+            td['std'] = self.std
+            td['text_clip_features'] = self.text_clip_features
+        if self.clip_video_required:
+            td['clip_features'] = self.clip_features
+        return td
+
+    def sample(self, idx: int) -> dict[str, torch.Tensor]:
+        video_id = self.videos[idx]
+        
+        if video_id in self.captions and torch.rand(1).item() < self.autoacd_sample_prob:
+            label = self.captions[video_id]
+        else:
+            label = self.labels[video_id]
+            
+        reader = StreamingMediaDecoder(self.root / (video_id + '.mp4'))
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_SYNC_FPS * self.duration_sec),
+            frame_rate=_SYNC_FPS,
+            format='rgb24',
+        )
+        if self.audio_required:
+            reader.add_basic_audio_stream(frames_per_chunk=2**30, )
+        if self.clip_video_required:
+            reader.add_basic_video_stream(
+                frames_per_chunk=int(_CLIP_FPS * self.duration_sec),
+                frame_rate=_CLIP_FPS,
+                format='rgb24',
+            )
+
+        reader.fill_buffer()
+        data_chunk = reader.pop_chunks()
+        
+        sync_chunk = data_chunk[0]
+        if sync_chunk is None:
+            raise RuntimeError(f'Sync video returned None {video_id}')
+        sync_chunk = normalize_video_chunk(sync_chunk, self.sync_expected_length, 
+                                           n_tolerance_frame=3, desc=video_id)
+        sync_chunk = self.sync_transform(sync_chunk)
+        
+        if self.audio_required:
+            audio_chunk = data_chunk[1]
+            
+        if self.clip_video_required:
+            clip_chunk = data_chunk[2 if self.audio_required else 1]
+            if clip_chunk is None:
+                raise RuntimeError(f'CLIP video returned None {video_id}')
+            clip_chunk = normalize_video_chunk(clip_chunk, self.clip_expected_length,
+                                               n_tolerance_frame=1, desc=video_id)
+            clip_chunk = self.clip_transform(clip_chunk)
+        
+        # process audio
+        if self.audio_required:
+            sample_rate = int(reader.get_out_stream_info(1).sample_rate)
+            audio_chunk = audio_chunk.transpose(0, 1)
+            audio_chunk = audio_chunk.mean(dim=0)  # mono
+            if self.normalize_audio:
+                abs_max = audio_chunk.abs().max()
+                audio_chunk = audio_chunk * (0.95 / abs_max)
+                if abs_max <= 1e-6:
+                    raise RuntimeError(f'Audio is silent {video_id}')
+
+            # resample
+            if sample_rate == self.sample_rate:
+                audio_chunk = audio_chunk
+            else:
+                if sample_rate not in self.resampler:
+                    # https://pytorch.org/audio/stable/tutorials/audio_resampling_tutorial.html#kaiser-best
+                    self.resampler[sample_rate] = torchaudio.transforms.Resample(
+                        sample_rate,
+                        self.sample_rate,
+                        lowpass_filter_width=64,
+                        rolloff=0.9475937167399596,
+                        resampling_method='sinc_interp_kaiser',
+                        beta=14.769656459379492,
+                    )
+                audio_chunk = self.resampler[sample_rate](audio_chunk)
+
+            if audio_chunk.shape[0] < self.expected_audio_length:
+                raise RuntimeError(f'Audio too short {video_id}')
+            audio_chunk = audio_chunk[:self.expected_audio_length]
+
+        data = {
+            'id': video_id,
+            'caption': label,
+            'sync_video': sync_chunk,
+            'sync_f_vid_orig': self.sync_features[self.id2idx_mmap[video_id]],
+            'text_features': self.text_features[self.id2idx_mmap_tsync[video_id]],
+            'text_masks': self.text_masks[self.id2idx_mmap_tsync[video_id]],
+            'video_exist': self.video_exist,
+            'text_exist': self.text_exist,
+        }
+        
+        if self.for_generator:
+            data['a_mean'] = self.mean[self.id2idx_mmap[video_id]]
+            data['a_std'] = self.std[self.id2idx_mmap[video_id]]
+            data['text_clip_features'] = self.text_clip_features[self.id2idx_mmap[video_id]]
+        
+        if self.audio_required:
+            data['audio'] = audio_chunk
+            
+        if self.clip_video_required:
+            data['clip_video'] = clip_chunk
+            data['clip_features'] = self.clip_features[self.id2idx_mmap[video_id]],
+
+        return data
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        try:
+            return self.sample(idx)
+        except Exception as e:
+            log.error(f'Error loading video {self.videos[idx]}: {e}')
+            return None
+
+    def __len__(self):
+        return len(self.labels)

selva_core/ext/__init__.py

@@ -0,0 +1 @@
+

selva_core/ext/autoencoder/__init__.py

@@ -0,0 +1 @@
+from .autoencoder import AutoEncoderModule

selva_core/ext/autoencoder/autoencoder.py

@@ -0,0 +1,52 @@
+from typing import Literal, Optional
+
+import torch
+import torch.nn as nn
+
+from selva_core.ext.autoencoder.vae import VAE, get_my_vae
+from selva_core.ext.bigvgan import BigVGAN
+from selva_core.ext.bigvgan_v2.bigvgan import BigVGAN as BigVGANv2
+from selva_core.model.utils.distributions import DiagonalGaussianDistribution
+
+
+class AutoEncoderModule(nn.Module):
+
+    def __init__(self,
+                 *,
+                 vae_ckpt_path,
+                 vocoder_ckpt_path: Optional[str] = None,
+                 mode: Literal['16k', '44k'],
+                 need_vae_encoder: bool = True):
+        super().__init__()
+        self.vae: VAE = get_my_vae(mode).eval()
+        vae_state_dict = torch.load(vae_ckpt_path, weights_only=True, map_location='cpu')
+        self.vae.load_state_dict(vae_state_dict)
+        self.vae.remove_weight_norm()
+
+        if mode == '16k':
+            assert vocoder_ckpt_path is not None
+            self.vocoder = BigVGAN(vocoder_ckpt_path).eval()
+        elif mode == '44k':
+            self.vocoder = BigVGANv2.from_pretrained('nvidia/bigvgan_v2_44khz_128band_512x',
+                                                     use_cuda_kernel=False)
+            self.vocoder.remove_weight_norm()
+        else:
+            raise ValueError(f'Unknown mode: {mode}')
+
+        for param in self.parameters():
+            param.requires_grad = False
+
+        if not need_vae_encoder:
+            del self.vae.encoder
+
+    @torch.inference_mode()
+    def encode(self, x: torch.Tensor) -> DiagonalGaussianDistribution:
+        return self.vae.encode(x)
+
+    @torch.inference_mode()
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        return self.vae.decode(z)
+
+    @torch.inference_mode()
+    def vocode(self, spec: torch.Tensor) -> torch.Tensor:
+        return self.vocoder(spec)

selva_core/ext/autoencoder/edm2_utils.py

@@ -0,0 +1,168 @@
+# Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# This work is licensed under a Creative Commons
+# Attribution-NonCommercial-ShareAlike 4.0 International License.
+# You should have received a copy of the license along with this
+# work. If not, see http://creativecommons.org/licenses/by-nc-sa/4.0/
+"""Improved diffusion model architecture proposed in the paper
+"Analyzing and Improving the Training Dynamics of Diffusion Models"."""
+
+import numpy as np
+import torch
+
+#----------------------------------------------------------------------------
+# Variant of constant() that inherits dtype and device from the given
+# reference tensor by default.
+
+_constant_cache = dict()
+
+
+def constant(value, shape=None, dtype=None, device=None, memory_format=None):
+    value = np.asarray(value)
+    if shape is not None:
+        shape = tuple(shape)
+    if dtype is None:
+        dtype = torch.get_default_dtype()
+    if device is None:
+        device = torch.device('cpu')
+    if memory_format is None:
+        memory_format = torch.contiguous_format
+
+    key = (value.shape, value.dtype, value.tobytes(), shape, dtype, device, memory_format)
+    tensor = _constant_cache.get(key, None)
+    if tensor is None:
+        tensor = torch.as_tensor(value.copy(), dtype=dtype, device=device)
+        if shape is not None:
+            tensor, _ = torch.broadcast_tensors(tensor, torch.empty(shape))
+        tensor = tensor.contiguous(memory_format=memory_format)
+        _constant_cache[key] = tensor
+    return tensor
+
+
+def const_like(ref, value, shape=None, dtype=None, device=None, memory_format=None):
+    if dtype is None:
+        dtype = ref.dtype
+    if device is None:
+        device = ref.device
+    return constant(value, shape=shape, dtype=dtype, device=device, memory_format=memory_format)
+
+
+#----------------------------------------------------------------------------
+# Normalize given tensor to unit magnitude with respect to the given
+# dimensions. Default = all dimensions except the first.
+
+
+def normalize(x, dim=None, eps=1e-4):
+    if dim is None:
+        dim = list(range(1, x.ndim))
+    norm = torch.linalg.vector_norm(x, dim=dim, keepdim=True, dtype=torch.float32)
+    norm = torch.add(eps, norm, alpha=np.sqrt(norm.numel() / x.numel()))
+    return x / norm.to(x.dtype)
+
+
+class Normalize(torch.nn.Module):
+
+    def __init__(self, dim=None, eps=1e-4):
+        super().__init__()
+        self.dim = dim
+        self.eps = eps
+
+    def forward(self, x):
+        return normalize(x, dim=self.dim, eps=self.eps)
+
+
+#----------------------------------------------------------------------------
+# Upsample or downsample the given tensor with the given filter,
+# or keep it as is.
+
+
+def resample(x, f=[1, 1], mode='keep'):
+    if mode == 'keep':
+        return x
+    f = np.float32(f)
+    assert f.ndim == 1 and len(f) % 2 == 0
+    pad = (len(f) - 1) // 2
+    f = f / f.sum()
+    f = np.outer(f, f)[np.newaxis, np.newaxis, :, :]
+    f = const_like(x, f)
+    c = x.shape[1]
+    if mode == 'down':
+        return torch.nn.functional.conv2d(x,
+                                          f.tile([c, 1, 1, 1]),
+                                          groups=c,
+                                          stride=2,
+                                          padding=(pad, ))
+    assert mode == 'up'
+    return torch.nn.functional.conv_transpose2d(x, (f * 4).tile([c, 1, 1, 1]),
+                                                groups=c,
+                                                stride=2,
+                                                padding=(pad, ))
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving SiLU (Equation 81).
+
+
+def mp_silu(x):
+    return torch.nn.functional.silu(x) / 0.596
+
+
+class MPSiLU(torch.nn.Module):
+
+    def forward(self, x):
+        return mp_silu(x)
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving sum (Equation 88).
+
+
+def mp_sum(a, b, t=0.5):
+    return a.lerp(b, t) / np.sqrt((1 - t)**2 + t**2)
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving concatenation (Equation 103).
+
+
+def mp_cat(a, b, dim=1, t=0.5):
+    Na = a.shape[dim]
+    Nb = b.shape[dim]
+    C = np.sqrt((Na + Nb) / ((1 - t)**2 + t**2))
+    wa = C / np.sqrt(Na) * (1 - t)
+    wb = C / np.sqrt(Nb) * t
+    return torch.cat([wa * a, wb * b], dim=dim)
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving convolution or fully-connected layer (Equation 47)
+# with force weight normalization (Equation 66).
+
+
+class MPConv1D(torch.nn.Module):
+
+    def __init__(self, in_channels, out_channels, kernel_size):
+        super().__init__()
+        self.out_channels = out_channels
+        self.weight = torch.nn.Parameter(torch.randn(out_channels, in_channels, kernel_size))
+
+        self.weight_norm_removed = False
+
+    def forward(self, x, gain=1):
+        assert self.weight_norm_removed, 'call remove_weight_norm() before inference'
+
+        w = self.weight * gain
+        if w.ndim == 2:
+            return x @ w.t()
+        assert w.ndim == 3
+        return torch.nn.functional.conv1d(x, w, padding=(w.shape[-1] // 2, ))
+
+    def remove_weight_norm(self):
+        w = self.weight.to(torch.float32)
+        w = normalize(w)  # traditional weight normalization
+        w = w / np.sqrt(w[0].numel())
+        w = w.to(self.weight.dtype)
+        self.weight.data.copy_(w)
+
+        self.weight_norm_removed = True
+        return self

selva_core/ext/autoencoder/vae.py

@@ -0,0 +1,369 @@
+import logging
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from selva_core.ext.autoencoder.edm2_utils import MPConv1D
+from selva_core.ext.autoencoder.vae_modules import (AttnBlock1D, Downsample1D, ResnetBlock1D,
+                                                 Upsample1D, nonlinearity)
+from selva_core.model.utils.distributions import DiagonalGaussianDistribution
+
+log = logging.getLogger()
+
+DATA_MEAN_80D = [
+    -1.6058, -1.3676, -1.2520, -1.2453, -1.2078, -1.2224, -1.2419, -1.2439, -1.2922, -1.2927,
+    -1.3170, -1.3543, -1.3401, -1.3836, -1.3907, -1.3912, -1.4313, -1.4152, -1.4527, -1.4728,
+    -1.4568, -1.5101, -1.5051, -1.5172, -1.5623, -1.5373, -1.5746, -1.5687, -1.6032, -1.6131,
+    -1.6081, -1.6331, -1.6489, -1.6489, -1.6700, -1.6738, -1.6953, -1.6969, -1.7048, -1.7280,
+    -1.7361, -1.7495, -1.7658, -1.7814, -1.7889, -1.8064, -1.8221, -1.8377, -1.8417, -1.8643,
+    -1.8857, -1.8929, -1.9173, -1.9379, -1.9531, -1.9673, -1.9824, -2.0042, -2.0215, -2.0436,
+    -2.0766, -2.1064, -2.1418, -2.1855, -2.2319, -2.2767, -2.3161, -2.3572, -2.3954, -2.4282,
+    -2.4659, -2.5072, -2.5552, -2.6074, -2.6584, -2.7107, -2.7634, -2.8266, -2.8981, -2.9673
+]
+
+DATA_STD_80D = [
+    1.0291, 1.0411, 1.0043, 0.9820, 0.9677, 0.9543, 0.9450, 0.9392, 0.9343, 0.9297, 0.9276, 0.9263,
+    0.9242, 0.9254, 0.9232, 0.9281, 0.9263, 0.9315, 0.9274, 0.9247, 0.9277, 0.9199, 0.9188, 0.9194,
+    0.9160, 0.9161, 0.9146, 0.9161, 0.9100, 0.9095, 0.9145, 0.9076, 0.9066, 0.9095, 0.9032, 0.9043,
+    0.9038, 0.9011, 0.9019, 0.9010, 0.8984, 0.8983, 0.8986, 0.8961, 0.8962, 0.8978, 0.8962, 0.8973,
+    0.8993, 0.8976, 0.8995, 0.9016, 0.8982, 0.8972, 0.8974, 0.8949, 0.8940, 0.8947, 0.8936, 0.8939,
+    0.8951, 0.8956, 0.9017, 0.9167, 0.9436, 0.9690, 1.0003, 1.0225, 1.0381, 1.0491, 1.0545, 1.0604,
+    1.0761, 1.0929, 1.1089, 1.1196, 1.1176, 1.1156, 1.1117, 1.1070
+]
+
+DATA_MEAN_128D = [
+    -3.3462, -2.6723, -2.4893, -2.3143, -2.2664, -2.3317, -2.1802, -2.4006, -2.2357, -2.4597,
+    -2.3717, -2.4690, -2.5142, -2.4919, -2.6610, -2.5047, -2.7483, -2.5926, -2.7462, -2.7033,
+    -2.7386, -2.8112, -2.7502, -2.9594, -2.7473, -3.0035, -2.8891, -2.9922, -2.9856, -3.0157,
+    -3.1191, -2.9893, -3.1718, -3.0745, -3.1879, -3.2310, -3.1424, -3.2296, -3.2791, -3.2782,
+    -3.2756, -3.3134, -3.3509, -3.3750, -3.3951, -3.3698, -3.4505, -3.4509, -3.5089, -3.4647,
+    -3.5536, -3.5788, -3.5867, -3.6036, -3.6400, -3.6747, -3.7072, -3.7279, -3.7283, -3.7795,
+    -3.8259, -3.8447, -3.8663, -3.9182, -3.9605, -3.9861, -4.0105, -4.0373, -4.0762, -4.1121,
+    -4.1488, -4.1874, -4.2461, -4.3170, -4.3639, -4.4452, -4.5282, -4.6297, -4.7019, -4.7960,
+    -4.8700, -4.9507, -5.0303, -5.0866, -5.1634, -5.2342, -5.3242, -5.4053, -5.4927, -5.5712,
+    -5.6464, -5.7052, -5.7619, -5.8410, -5.9188, -6.0103, -6.0955, -6.1673, -6.2362, -6.3120,
+    -6.3926, -6.4797, -6.5565, -6.6511, -6.8130, -6.9961, -7.1275, -7.2457, -7.3576, -7.4663,
+    -7.6136, -7.7469, -7.8815, -8.0132, -8.1515, -8.3071, -8.4722, -8.7418, -9.3975, -9.6628,
+    -9.7671, -9.8863, -9.9992, -10.0860, -10.1709, -10.5418, -11.2795, -11.3861
+]
+
+DATA_STD_128D = [
+    2.3804, 2.4368, 2.3772, 2.3145, 2.2803, 2.2510, 2.2316, 2.2083, 2.1996, 2.1835, 2.1769, 2.1659,
+    2.1631, 2.1618, 2.1540, 2.1606, 2.1571, 2.1567, 2.1612, 2.1579, 2.1679, 2.1683, 2.1634, 2.1557,
+    2.1668, 2.1518, 2.1415, 2.1449, 2.1406, 2.1350, 2.1313, 2.1415, 2.1281, 2.1352, 2.1219, 2.1182,
+    2.1327, 2.1195, 2.1137, 2.1080, 2.1179, 2.1036, 2.1087, 2.1036, 2.1015, 2.1068, 2.0975, 2.0991,
+    2.0902, 2.1015, 2.0857, 2.0920, 2.0893, 2.0897, 2.0910, 2.0881, 2.0925, 2.0873, 2.0960, 2.0900,
+    2.0957, 2.0958, 2.0978, 2.0936, 2.0886, 2.0905, 2.0845, 2.0855, 2.0796, 2.0840, 2.0813, 2.0817,
+    2.0838, 2.0840, 2.0917, 2.1061, 2.1431, 2.1976, 2.2482, 2.3055, 2.3700, 2.4088, 2.4372, 2.4609,
+    2.4731, 2.4847, 2.5072, 2.5451, 2.5772, 2.6147, 2.6529, 2.6596, 2.6645, 2.6726, 2.6803, 2.6812,
+    2.6899, 2.6916, 2.6931, 2.6998, 2.7062, 2.7262, 2.7222, 2.7158, 2.7041, 2.7485, 2.7491, 2.7451,
+    2.7485, 2.7233, 2.7297, 2.7233, 2.7145, 2.6958, 2.6788, 2.6439, 2.6007, 2.4786, 2.2469, 2.1877,
+    2.1392, 2.0717, 2.0107, 1.9676, 1.9140, 1.7102, 0.9101, 0.7164
+]
+
+
+class VAE(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        data_dim: int,
+        embed_dim: int,
+        hidden_dim: int,
+    ):
+        super().__init__()
+
+        if data_dim == 80:
+            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_80D, dtype=torch.float32))
+            self.data_std = nn.Buffer(torch.tensor(DATA_STD_80D, dtype=torch.float32))
+        elif data_dim == 128:
+            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_128D, dtype=torch.float32))
+            self.data_std = nn.Buffer(torch.tensor(DATA_STD_128D, dtype=torch.float32))
+
+        self.data_mean = self.data_mean.view(1, -1, 1)
+        self.data_std = self.data_std.view(1, -1, 1)
+
+        self.encoder = Encoder1D(
+            dim=hidden_dim,
+            ch_mult=(1, 2, 4),
+            num_res_blocks=2,
+            attn_layers=[3],
+            down_layers=[0],
+            in_dim=data_dim,
+            embed_dim=embed_dim,
+        )
+        self.decoder = Decoder1D(
+            dim=hidden_dim,
+            ch_mult=(1, 2, 4),
+            num_res_blocks=2,
+            attn_layers=[3],
+            down_layers=[0],
+            in_dim=data_dim,
+            out_dim=data_dim,
+            embed_dim=embed_dim,
+        )
+
+        self.embed_dim = embed_dim
+        # self.quant_conv = nn.Conv1d(2 * embed_dim, 2 * embed_dim, 1)
+        # self.post_quant_conv = nn.Conv1d(embed_dim, embed_dim, 1)
+
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        pass
+
+    def encode(self, x: torch.Tensor, normalize: bool = True) -> DiagonalGaussianDistribution:
+        if normalize:
+            x = self.normalize(x)
+        moments = self.encoder(x)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z: torch.Tensor, unnormalize: bool = True) -> torch.Tensor:
+        dec = self.decoder(z)
+        if unnormalize:
+            dec = self.unnormalize(dec)
+        return dec
+
+    def normalize(self, x: torch.Tensor) -> torch.Tensor:
+        return (x - self.data_mean) / self.data_std
+
+    def unnormalize(self, x: torch.Tensor) -> torch.Tensor:
+        return x * self.data_std + self.data_mean
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        sample_posterior: bool = True,
+        rng: Optional[torch.Generator] = None,
+        normalize: bool = True,
+        unnormalize: bool = True,
+    ) -> tuple[torch.Tensor, DiagonalGaussianDistribution]:
+
+        posterior = self.encode(x, normalize=normalize)
+        if sample_posterior:
+            z = posterior.sample(rng)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, unnormalize=unnormalize)
+        return dec, posterior
+
+    def load_weights(self, src_dict) -> None:
+        self.load_state_dict(src_dict, strict=True)
+
+    @property
+    def device(self) -> torch.device:
+        return next(self.parameters()).device
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def remove_weight_norm(self):
+        for name, m in self.named_modules():
+            if isinstance(m, MPConv1D):
+                m.remove_weight_norm()
+                log.debug(f"Removed weight norm from {name}")
+        return self
+
+
+class Encoder1D(nn.Module):
+
+    def __init__(self,
+                 *,
+                 dim: int,
+                 ch_mult: tuple[int] = (1, 2, 4, 8),
+                 num_res_blocks: int,
+                 attn_layers: list[int] = [],
+                 down_layers: list[int] = [],
+                 resamp_with_conv: bool = True,
+                 in_dim: int,
+                 embed_dim: int,
+                 double_z: bool = True,
+                 kernel_size: int = 3,
+                 clip_act: float = 256.0):
+        super().__init__()
+        self.dim = dim
+        self.num_layers = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.in_channels = in_dim
+        self.clip_act = clip_act
+        self.down_layers = down_layers
+        self.attn_layers = attn_layers
+        self.conv_in = MPConv1D(in_dim, self.dim, kernel_size=kernel_size)
+
+        in_ch_mult = (1, ) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        # downsampling
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_layers):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = dim * in_ch_mult[i_level]
+            block_out = dim * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock1D(in_dim=block_in,
+                                  out_dim=block_out,
+                                  kernel_size=kernel_size,
+                                  use_norm=True))
+                block_in = block_out
+                if i_level in attn_layers:
+                    attn.append(AttnBlock1D(block_in))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level in down_layers:
+                down.downsample = Downsample1D(block_in, resamp_with_conv)
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock1D(in_dim=block_in,
+                                         out_dim=block_in,
+                                         kernel_size=kernel_size,
+                                         use_norm=True)
+        self.mid.attn_1 = AttnBlock1D(block_in)
+        self.mid.block_2 = ResnetBlock1D(in_dim=block_in,
+                                         out_dim=block_in,
+                                         kernel_size=kernel_size,
+                                         use_norm=True)
+
+        # end
+        self.conv_out = MPConv1D(block_in,
+                                 2 * embed_dim if double_z else embed_dim,
+                                 kernel_size=kernel_size)
+
+        self.learnable_gain = nn.Parameter(torch.zeros([]))
+
+    def forward(self, x):
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_layers):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1])
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                h = h.clamp(-self.clip_act, self.clip_act)
+                hs.append(h)
+            if i_level in self.down_layers:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        h = h.clamp(-self.clip_act, self.clip_act)
+
+        # end
+        h = nonlinearity(h)
+        h = self.conv_out(h, gain=(self.learnable_gain + 1))
+        return h
+
+
+class Decoder1D(nn.Module):
+
+    def __init__(self,
+                 *,
+                 dim: int,
+                 out_dim: int,
+                 ch_mult: tuple[int] = (1, 2, 4, 8),
+                 num_res_blocks: int,
+                 attn_layers: list[int] = [],
+                 down_layers: list[int] = [],
+                 kernel_size: int = 3,
+                 resamp_with_conv: bool = True,
+                 in_dim: int,
+                 embed_dim: int,
+                 clip_act: float = 256.0):
+        super().__init__()
+        self.ch = dim
+        self.num_layers = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.in_channels = in_dim
+        self.clip_act = clip_act
+        self.down_layers = [i + 1 for i in down_layers]  # each downlayer add one
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = dim * ch_mult[self.num_layers - 1]
+
+        # z to block_in
+        self.conv_in = MPConv1D(embed_dim, block_in, kernel_size=kernel_size)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
+        self.mid.attn_1 = AttnBlock1D(block_in)
+        self.mid.block_2 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_layers)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = dim * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(ResnetBlock1D(in_dim=block_in, out_dim=block_out, use_norm=True))
+                block_in = block_out
+                if i_level in attn_layers:
+                    attn.append(AttnBlock1D(block_in))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level in self.down_layers:
+                up.upsample = Upsample1D(block_in, resamp_with_conv)
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.conv_out = MPConv1D(block_in, out_dim, kernel_size=kernel_size)
+        self.learnable_gain = nn.Parameter(torch.zeros([]))
+
+    def forward(self, z):
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        h = h.clamp(-self.clip_act, self.clip_act)
+
+        # upsampling
+        for i_level in reversed(range(self.num_layers)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+                h = h.clamp(-self.clip_act, self.clip_act)
+            if i_level in self.down_layers:
+                h = self.up[i_level].upsample(h)
+
+        h = nonlinearity(h)
+        h = self.conv_out(h, gain=(self.learnable_gain + 1))
+        return h
+
+
+def VAE_16k(**kwargs) -> VAE:
+    return VAE(data_dim=80, embed_dim=20, hidden_dim=384, **kwargs)
+
+
+def VAE_44k(**kwargs) -> VAE:
+    return VAE(data_dim=128, embed_dim=40, hidden_dim=512, **kwargs)
+
+
+def get_my_vae(name: str, **kwargs) -> VAE:
+    if name == '16k':
+        return VAE_16k(**kwargs)
+    if name == '44k':
+        return VAE_44k(**kwargs)
+    raise ValueError(f'Unknown model: {name}')
+
+
+if __name__ == '__main__':
+    network = get_my_vae('standard')
+
+    # print the number of parameters in terms of millions
+    num_params = sum(p.numel() for p in network.parameters()) / 1e6
+    print(f'Number of parameters: {num_params:.2f}M')

selva_core/ext/autoencoder/vae_modules.py

@@ -0,0 +1,117 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from selva_core.ext.autoencoder.edm2_utils import (MPConv1D, mp_silu, mp_sum, normalize)
+
+
+def nonlinearity(x):
+    # swish
+    return mp_silu(x)
+
+
+class ResnetBlock1D(nn.Module):
+
+    def __init__(self, *, in_dim, out_dim=None, conv_shortcut=False, kernel_size=3, use_norm=True):
+        super().__init__()
+        self.in_dim = in_dim
+        out_dim = in_dim if out_dim is None else out_dim
+        self.out_dim = out_dim
+        self.use_conv_shortcut = conv_shortcut
+        self.use_norm = use_norm
+
+        self.conv1 = MPConv1D(in_dim, out_dim, kernel_size=kernel_size)
+        self.conv2 = MPConv1D(out_dim, out_dim, kernel_size=kernel_size)
+        if self.in_dim != self.out_dim:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = MPConv1D(in_dim, out_dim, kernel_size=kernel_size)
+            else:
+                self.nin_shortcut = MPConv1D(in_dim, out_dim, kernel_size=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+
+        # pixel norm
+        if self.use_norm:
+            x = normalize(x, dim=1)
+
+        h = x
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        h = nonlinearity(h)
+        h = self.conv2(h)
+
+        if self.in_dim != self.out_dim:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return mp_sum(x, h, t=0.3)
+
+
+class AttnBlock1D(nn.Module):
+
+    def __init__(self, in_channels, num_heads=1):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.num_heads = num_heads
+        self.qkv = MPConv1D(in_channels, in_channels * 3, kernel_size=1)
+        self.proj_out = MPConv1D(in_channels, in_channels, kernel_size=1)
+
+    def forward(self, x):
+        h = x
+        y = self.qkv(h)
+        y = y.reshape(y.shape[0], self.num_heads, -1, 3, y.shape[-1])
+        q, k, v = normalize(y, dim=2).unbind(3)
+
+        q = rearrange(q, 'b h c l -> b h l c')
+        k = rearrange(k, 'b h c l -> b h l c')
+        v = rearrange(v, 'b h c l -> b h l c')
+
+        h = F.scaled_dot_product_attention(q, k, v)
+        h = rearrange(h, 'b h l c -> b (h c) l')
+
+        h = self.proj_out(h)
+
+        return mp_sum(x, h, t=0.3)
+
+
+class Upsample1D(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = MPConv1D(in_channels, in_channels, kernel_size=3)
+
+    def forward(self, x):
+        x = F.interpolate(x, scale_factor=2.0, mode='nearest-exact')  # support 3D tensor(B,C,T)
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample1D(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv1 = MPConv1D(in_channels, in_channels, kernel_size=1)
+            self.conv2 = MPConv1D(in_channels, in_channels, kernel_size=1)
+
+    def forward(self, x):
+
+        if self.with_conv:
+            x = self.conv1(x)
+
+        x = F.avg_pool1d(x, kernel_size=2, stride=2)
+
+        if self.with_conv:
+            x = self.conv2(x)
+
+        return x

selva_core/ext/bigvgan/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 NVIDIA CORPORATION.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software. 
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan/__init__.py

@@ -0,0 +1 @@
+from .bigvgan import BigVGAN

selva_core/ext/bigvgan/activations.py

@@ -0,0 +1,120 @@
+# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+from torch import nn, sin, pow
+from torch.nn import Parameter
+
+
+class Snake(nn.Module):
+    '''
+    Implementation of a sine-based periodic activation function
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter
+    References:
+        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snake(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    '''
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
+        '''
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha: trainable parameter
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            alpha will be trained along with the rest of your model.
+        '''
+        super(Snake, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale: # log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+        else: # linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        '''
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        Snake ∶= x + 1/a * sin^2 (xa)
+        '''
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x
+
+
+class SnakeBeta(nn.Module):
+    '''
+    A modified Snake function which uses separate parameters for the magnitude of the periodic components
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter that controls frequency
+        - beta - trainable parameter that controls magnitude
+    References:
+        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snakebeta(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    '''
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
+        '''
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha - trainable parameter that controls frequency
+            - beta - trainable parameter that controls magnitude
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            beta is initialized to 1 by default, higher values = higher-magnitude.
+            alpha will be trained along with the rest of your model.
+        '''
+        super(SnakeBeta, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale: # log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+            self.beta = Parameter(torch.zeros(in_features) * alpha)
+        else: # linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+            self.beta = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        '''
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        SnakeBeta ∶= x + 1/b * sin^2 (xa)
+        '''
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
+        beta = self.beta.unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+            beta = torch.exp(beta)
+        x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x

selva_core/ext/bigvgan/alias_free_torch/__init__.py

@@ -0,0 +1,6 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+from .filter import *
+from .resample import *
+from .act import *

selva_core/ext/bigvgan/alias_free_torch/act.py

@@ -0,0 +1,28 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch.nn as nn
+from .resample import UpSample1d, DownSample1d
+
+
+class Activation1d(nn.Module):
+    def __init__(self,
+                 activation,
+                 up_ratio: int = 2,
+                 down_ratio: int = 2,
+                 up_kernel_size: int = 12,
+                 down_kernel_size: int = 12):
+        super().__init__()
+        self.up_ratio = up_ratio
+        self.down_ratio = down_ratio
+        self.act = activation
+        self.upsample = UpSample1d(up_ratio, up_kernel_size)
+        self.downsample = DownSample1d(down_ratio, down_kernel_size)
+
+    # x: [B,C,T]
+    def forward(self, x):
+        x = self.upsample(x)
+        x = self.act(x)
+        x = self.downsample(x)
+
+        return x

selva_core/ext/bigvgan/alias_free_torch/filter.py

@@ -0,0 +1,95 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+if 'sinc' in dir(torch):
+    sinc = torch.sinc
+else:
+    # This code is adopted from adefossez's julius.core.sinc under the MIT License
+    # https://adefossez.github.io/julius/julius/core.html
+    #   LICENSE is in incl_licenses directory.
+    def sinc(x: torch.Tensor):
+        """
+        Implementation of sinc, i.e. sin(pi * x) / (pi * x)
+        __Warning__: Different to julius.sinc, the input is multiplied by `pi`!
+        """
+        return torch.where(x == 0,
+                           torch.tensor(1., device=x.device, dtype=x.dtype),
+                           torch.sin(math.pi * x) / math.pi / x)
+
+
+# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License
+# https://adefossez.github.io/julius/julius/lowpass.html
+#   LICENSE is in incl_licenses directory.
+def kaiser_sinc_filter1d(cutoff, half_width, kernel_size): # return filter [1,1,kernel_size]
+    even = (kernel_size % 2 == 0)
+    half_size = kernel_size // 2
+
+    #For kaiser window
+    delta_f = 4 * half_width
+    A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95
+    if A > 50.:
+        beta = 0.1102 * (A - 8.7)
+    elif A >= 21.:
+        beta = 0.5842 * (A - 21)**0.4 + 0.07886 * (A - 21.)
+    else:
+        beta = 0.
+    window = torch.kaiser_window(kernel_size, beta=beta, periodic=False)
+
+    # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio
+    if even:
+        time = (torch.arange(-half_size, half_size) + 0.5)
+    else:
+        time = torch.arange(kernel_size) - half_size
+    if cutoff == 0:
+        filter_ = torch.zeros_like(time)
+    else:
+        filter_ = 2 * cutoff * window * sinc(2 * cutoff * time)
+        # Normalize filter to have sum = 1, otherwise we will have a small leakage
+        # of the constant component in the input signal.
+        filter_ /= filter_.sum()
+        filter = filter_.view(1, 1, kernel_size)
+
+    return filter
+
+
+class LowPassFilter1d(nn.Module):
+    def __init__(self,
+                 cutoff=0.5,
+                 half_width=0.6,
+                 stride: int = 1,
+                 padding: bool = True,
+                 padding_mode: str = 'replicate',
+                 kernel_size: int = 12):
+        # kernel_size should be even number for stylegan3 setup,
+        # in this implementation, odd number is also possible.
+        super().__init__()
+        if cutoff < -0.:
+            raise ValueError("Minimum cutoff must be larger than zero.")
+        if cutoff > 0.5:
+            raise ValueError("A cutoff above 0.5 does not make sense.")
+        self.kernel_size = kernel_size
+        self.even = (kernel_size % 2 == 0)
+        self.pad_left = kernel_size // 2 - int(self.even)
+        self.pad_right = kernel_size // 2
+        self.stride = stride
+        self.padding = padding
+        self.padding_mode = padding_mode
+        filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size)
+        self.register_buffer("filter", filter)
+
+    #input [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        if self.padding:
+            x = F.pad(x, (self.pad_left, self.pad_right),
+                      mode=self.padding_mode)
+        out = F.conv1d(x, self.filter.expand(C, -1, -1),
+                       stride=self.stride, groups=C)
+
+        return out

selva_core/ext/bigvgan/alias_free_torch/resample.py

@@ -0,0 +1,49 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch.nn as nn
+from torch.nn import functional as F
+from .filter import LowPassFilter1d
+from .filter import kaiser_sinc_filter1d
+
+
+class UpSample1d(nn.Module):
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+        self.stride = ratio
+        self.pad = self.kernel_size // ratio - 1
+        self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2
+        self.pad_right = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2
+        filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio,
+                                      half_width=0.6 / ratio,
+                                      kernel_size=self.kernel_size)
+        self.register_buffer("filter", filter)
+
+    # x: [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        x = F.pad(x, (self.pad, self.pad), mode='replicate')
+        x = self.ratio * F.conv_transpose1d(
+            x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C)
+        x = x[..., self.pad_left:-self.pad_right]
+
+        return x
+
+
+class DownSample1d(nn.Module):
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+        self.lowpass = LowPassFilter1d(cutoff=0.5 / ratio,
+                                       half_width=0.6 / ratio,
+                                       stride=ratio,
+                                       kernel_size=self.kernel_size)
+
+    def forward(self, x):
+        xx = self.lowpass(x)
+
+        return xx

selva_core/ext/bigvgan/bigvgan.py

@@ -0,0 +1,32 @@
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from omegaconf import OmegaConf
+
+from selva_core.ext.bigvgan.models import BigVGANVocoder
+
+_bigvgan_vocoder_path = Path(__file__).parent / 'bigvgan_vocoder.yml'
+
+
+class BigVGAN(nn.Module):
+
+    def __init__(self, ckpt_path, config_path=_bigvgan_vocoder_path):
+        super().__init__()
+        vocoder_cfg = OmegaConf.load(config_path)
+        self.vocoder = BigVGANVocoder(vocoder_cfg).eval()
+        vocoder_ckpt = torch.load(ckpt_path, map_location='cpu', weights_only=True)['generator']
+        self.vocoder.load_state_dict(vocoder_ckpt)
+
+        self.weight_norm_removed = False
+        self.remove_weight_norm()
+
+    @torch.inference_mode()
+    def forward(self, x):
+        assert self.weight_norm_removed, 'call remove_weight_norm() before inference'
+        return self.vocoder(x)
+
+    def remove_weight_norm(self):
+        self.vocoder.remove_weight_norm()
+        self.weight_norm_removed = True
+        return self

selva_core/ext/bigvgan/bigvgan_vocoder.yml

@@ -0,0 +1,63 @@
+resblock: '1'
+num_gpus: 0
+batch_size: 64
+num_mels: 80
+learning_rate: 0.0001
+adam_b1: 0.8
+adam_b2: 0.99
+lr_decay: 0.999
+seed: 1234
+upsample_rates:
+- 4
+- 4
+- 2
+- 2
+- 2
+- 2
+upsample_kernel_sizes:
+- 8
+- 8
+- 4
+- 4
+- 4
+- 4
+upsample_initial_channel: 1536
+resblock_kernel_sizes:
+- 3
+- 7
+- 11
+resblock_dilation_sizes:
+- - 1
+  - 3
+  - 5
+- - 1
+  - 3
+  - 5
+- - 1
+  - 3
+  - 5
+activation: snakebeta
+snake_logscale: true
+resolutions:
+- - 1024
+  - 120
+  - 600
+- - 2048
+  - 240
+  - 1200
+- - 512
+  - 50
+  - 240
+mpd_reshapes:
+- 2
+- 3
+- 5
+- 7
+- 11
+use_spectral_norm: false
+discriminator_channel_mult: 1
+num_workers: 4
+dist_config:
+  dist_backend: nccl
+  dist_url: tcp://localhost:54341
+  world_size: 1

selva_core/ext/bigvgan/env.py

@@ -0,0 +1,18 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import os
+import shutil
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+def build_env(config, config_name, path):
+    t_path = os.path.join(path, config_name)
+    if config != t_path:
+        os.makedirs(path, exist_ok=True)
+        shutil.copyfile(config, os.path.join(path, config_name))

selva_core/ext/bigvgan/incl_licenses/LICENSE_1

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Jungil Kong
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan/incl_licenses/LICENSE_2

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Edward Dixon
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan/incl_licenses/LICENSE_3

@@ -0,0 +1,201 @@
+                                Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

selva_core/ext/bigvgan/incl_licenses/LICENSE_4

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2019, Seungwon Park 박승원
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

selva_core/ext/bigvgan/incl_licenses/LICENSE_5

@@ -0,0 +1,16 @@
+Copyright 2020 Alexandre Défossez
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
+associated documentation files (the "Software"), to deal in the Software without restriction,
+including without limitation the rights to use, copy, modify, merge, publish, distribute,
+sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or
+substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT
+NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
+DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

selva_core/ext/bigvgan/models.py

@@ -0,0 +1,255 @@
+# Copyright (c) 2022 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+import torch.nn as nn
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils.parametrizations import weight_norm
+from torch.nn.utils.parametrize import remove_parametrizations
+
+from selva_core.ext.bigvgan import activations
+from selva_core.ext.bigvgan.alias_free_torch import *
+from selva_core.ext.bigvgan.utils import get_padding, init_weights
+
+LRELU_SLOPE = 0.1
+
+
+class AMPBlock1(torch.nn.Module):
+
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5), activation=None):
+        super(AMPBlock1, self).__init__()
+        self.h = h
+
+        self.convs1 = nn.ModuleList([
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[0],
+                       padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[1],
+                       padding=get_padding(kernel_size, dilation[1]))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[2],
+                       padding=get_padding(kernel_size, dilation[2])))
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList([
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1)))
+        ])
+        self.convs2.apply(init_weights)
+
+        self.num_layers = len(self.convs1) + len(self.convs2)  # total number of conv layers
+
+        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        acts1, acts2 = self.activations[::2], self.activations[1::2]
+        for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2):
+            xt = a1(x)
+            xt = c1(xt)
+            xt = a2(xt)
+            xt = c2(xt)
+            x = xt + x
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_parametrizations(l, 'weight')
+        for l in self.convs2:
+            remove_parametrizations(l, 'weight')
+
+
+class AMPBlock2(torch.nn.Module):
+
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3), activation=None):
+        super(AMPBlock2, self).__init__()
+        self.h = h
+
+        self.convs = nn.ModuleList([
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[0],
+                       padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[1],
+                       padding=get_padding(kernel_size, dilation[1])))
+        ])
+        self.convs.apply(init_weights)
+
+        self.num_layers = len(self.convs)  # total number of conv layers
+
+        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        for c, a in zip(self.convs, self.activations):
+            xt = a(x)
+            xt = c(xt)
+            x = xt + x
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_parametrizations(l, 'weight')
+
+
+class BigVGANVocoder(torch.nn.Module):
+    # this is our main BigVGAN model. Applies anti-aliased periodic activation for resblocks.
+    def __init__(self, h):
+        super().__init__()
+        self.h = h
+
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+
+        # pre conv
+        self.conv_pre = weight_norm(Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3))
+
+        # define which AMPBlock to use. BigVGAN uses AMPBlock1 as default
+        resblock = AMPBlock1 if h.resblock == '1' else AMPBlock2
+
+        # transposed conv-based upsamplers. does not apply anti-aliasing
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                nn.ModuleList([
+                    weight_norm(
+                        ConvTranspose1d(h.upsample_initial_channel // (2**i),
+                                        h.upsample_initial_channel // (2**(i + 1)),
+                                        k,
+                                        u,
+                                        padding=(k - u) // 2))
+                ]))
+
+        # residual blocks using anti-aliased multi-periodicity composition modules (AMP)
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2**(i + 1))
+            for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
+                self.resblocks.append(resblock(h, ch, k, d, activation=h.activation))
+
+        # post conv
+        if h.activation == "snake":  # periodic nonlinearity with snake function and anti-aliasing
+            activation_post = activations.Snake(ch, alpha_logscale=h.snake_logscale)
+            self.activation_post = Activation1d(activation=activation_post)
+        elif h.activation == "snakebeta":  # periodic nonlinearity with snakebeta function and anti-aliasing
+            activation_post = activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale)
+            self.activation_post = Activation1d(activation=activation_post)
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+
+        # weight initialization
+        for i in range(len(self.ups)):
+            self.ups[i].apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        # pre conv
+        x = self.conv_pre(x)
+
+        for i in range(self.num_upsamples):
+            # upsampling
+            for i_up in range(len(self.ups[i])):
+                x = self.ups[i][i_up](x)
+            # AMP blocks
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        # post conv
+        x = self.activation_post(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            for l_i in l:
+                remove_parametrizations(l_i, 'weight')
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_parametrizations(self.conv_pre, 'weight')
+        remove_parametrizations(self.conv_post, 'weight')

selva_core/ext/bigvgan/utils.py

@@ -0,0 +1,31 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import os
+
+import torch
+from torch.nn.utils.parametrizations import weight_norm
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def apply_weight_norm(m):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        weight_norm(m)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict

selva_core/ext/bigvgan_v2/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 NVIDIA CORPORATION.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan_v2/activations.py

@@ -0,0 +1,126 @@
+# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+from torch import nn, sin, pow
+from torch.nn import Parameter
+
+
+class Snake(nn.Module):
+    """
+    Implementation of a sine-based periodic activation function
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter
+    References:
+        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snake(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    """
+
+    def __init__(
+        self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False
+    ):
+        """
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha: trainable parameter
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            alpha will be trained along with the rest of your model.
+        """
+        super(Snake, self).__init__()
+        self.in_features = in_features
+
+        # Initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # Log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+        else:  # Linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        """
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        Snake ∶= x + 1/a * sin^2 (xa)
+        """
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # Line up with x to [B, C, T]
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x
+
+
+class SnakeBeta(nn.Module):
+    """
+    A modified Snake function which uses separate parameters for the magnitude of the periodic components
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter that controls frequency
+        - beta - trainable parameter that controls magnitude
+    References:
+        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snakebeta(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    """
+
+    def __init__(
+        self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False
+    ):
+        """
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha - trainable parameter that controls frequency
+            - beta - trainable parameter that controls magnitude
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            beta is initialized to 1 by default, higher values = higher-magnitude.
+            alpha will be trained along with the rest of your model.
+        """
+        super(SnakeBeta, self).__init__()
+        self.in_features = in_features
+
+        # Initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # Log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+            self.beta = Parameter(torch.zeros(in_features) * alpha)
+        else:  # Linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+            self.beta = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        """
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        SnakeBeta ∶= x + 1/b * sin^2 (xa)
+        """
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # Line up with x to [B, C, T]
+        beta = self.beta.unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+            beta = torch.exp(beta)
+        x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x

selva_core/ext/bigvgan_v2/alias_free_activation/cuda/activation1d.py

@@ -0,0 +1,77 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+import torch
+import torch.nn as nn
+from alias_free_activation.torch.resample import UpSample1d, DownSample1d
+
+# load fused CUDA kernel: this enables importing anti_alias_activation_cuda
+from alias_free_activation.cuda import load
+
+anti_alias_activation_cuda = load.load()
+
+
+class FusedAntiAliasActivation(torch.autograd.Function):
+    """
+    Assumes filter size 12, replication padding on upsampling/downsampling, and logscale alpha/beta parameters as inputs.
+    The hyperparameters are hard-coded in the kernel to maximize speed.
+    NOTE: The fused kenrel is incorrect for Activation1d with different hyperparameters.
+    """
+
+    @staticmethod
+    def forward(ctx, inputs, up_ftr, down_ftr, alpha, beta):
+        activation_results = anti_alias_activation_cuda.forward(
+            inputs, up_ftr, down_ftr, alpha, beta
+        )
+
+        return activation_results
+
+    @staticmethod
+    def backward(ctx, output_grads):
+        raise NotImplementedError
+        return output_grads, None, None
+
+
+class Activation1d(nn.Module):
+    def __init__(
+        self,
+        activation,
+        up_ratio: int = 2,
+        down_ratio: int = 2,
+        up_kernel_size: int = 12,
+        down_kernel_size: int = 12,
+        fused: bool = True,
+    ):
+        super().__init__()
+        self.up_ratio = up_ratio
+        self.down_ratio = down_ratio
+        self.act = activation
+        self.upsample = UpSample1d(up_ratio, up_kernel_size)
+        self.downsample = DownSample1d(down_ratio, down_kernel_size)
+
+        self.fused = fused  # Whether to use fused CUDA kernel or not
+
+    def forward(self, x):
+        if not self.fused:
+            x = self.upsample(x)
+            x = self.act(x)
+            x = self.downsample(x)
+            return x
+        else:
+            if self.act.__class__.__name__ == "Snake":
+                beta = self.act.alpha.data  # Snake uses same params for alpha and beta
+            else:
+                beta = (
+                    self.act.beta.data
+                )  # Snakebeta uses different params for alpha and beta
+            alpha = self.act.alpha.data
+            if (
+                not self.act.alpha_logscale
+            ):  # Exp baked into cuda kernel, cancel it out with a log
+                alpha = torch.log(alpha)
+                beta = torch.log(beta)
+
+            x = FusedAntiAliasActivation.apply(
+                x, self.upsample.filter, self.downsample.lowpass.filter, alpha, beta
+            )
+            return x

selva_core/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation.cpp

@@ -0,0 +1,23 @@
+/* coding=utf-8
+ * Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ #include <torch/extension.h>
+
+extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward", &fwd_cuda, "Anti-Alias Activation forward (CUDA)");
+}

selva_core/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation_cuda.cu

@@ -0,0 +1,246 @@
+/* coding=utf-8
+ * Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <ATen/ATen.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+#include <cuda_profiler_api.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <torch/extension.h>
+#include "type_shim.h"
+#include <assert.h>
+#include <cfloat>
+#include <limits>
+#include <stdint.h>
+#include <c10/macros/Macros.h>
+
+namespace
+{
+    // Hard-coded hyperparameters
+    // WARP_SIZE and WARP_BATCH must match the return values batches_per_warp and
+    constexpr int ELEMENTS_PER_LDG_STG = 1; //(WARP_ITERATIONS < 4) ? 1 : 4;
+    constexpr int BUFFER_SIZE = 32;
+    constexpr int FILTER_SIZE = 12;
+    constexpr int HALF_FILTER_SIZE = 6;
+    constexpr int UPSAMPLE_REPLICATION_PAD = 5; // 5 on each side, matching torch impl
+    constexpr int DOWNSAMPLE_REPLICATION_PAD_LEFT = 5; // matching torch impl
+    constexpr int DOWNSAMPLE_REPLICATION_PAD_RIGHT = 6; // matching torch impl
+
+    template <typename input_t, typename output_t, typename acc_t>
+    __global__ void anti_alias_activation_forward(
+        output_t *dst,
+        const input_t *src,
+        const input_t *up_ftr,
+        const input_t *down_ftr,
+        const input_t *alpha,
+        const input_t *beta,
+        int batch_size,
+        int channels,
+        int seq_len)
+    {
+        // Up and downsample filters
+        input_t up_filter[FILTER_SIZE];
+        input_t down_filter[FILTER_SIZE];
+
+        // Load data from global memory including extra indices reserved for replication paddings
+        input_t elements[2 * FILTER_SIZE + 2 * BUFFER_SIZE + 2 * UPSAMPLE_REPLICATION_PAD] = {0};
+        input_t intermediates[2 * FILTER_SIZE + 2 * BUFFER_SIZE + DOWNSAMPLE_REPLICATION_PAD_LEFT + DOWNSAMPLE_REPLICATION_PAD_RIGHT] = {0};
+
+        // Output stores downsampled output before writing to dst
+        output_t output[BUFFER_SIZE];
+
+        // blockDim/threadIdx = (128, 1, 1)
+        // gridDim/blockIdx = (seq_blocks, channels, batches)
+        int block_offset = (blockIdx.x * 128 * BUFFER_SIZE + seq_len * (blockIdx.y + gridDim.y * blockIdx.z));
+        int local_offset = threadIdx.x * BUFFER_SIZE;
+        int seq_offset = blockIdx.x * 128 * BUFFER_SIZE + local_offset;
+
+        // intermediate have double the seq_len
+        int intermediate_local_offset = threadIdx.x * BUFFER_SIZE * 2;
+        int intermediate_seq_offset = blockIdx.x * 128 * BUFFER_SIZE * 2 + intermediate_local_offset;
+
+        // Get values needed for replication padding before moving pointer
+        const input_t *right_most_pntr = src + (seq_len * (blockIdx.y + gridDim.y * blockIdx.z));
+        input_t seq_left_most_value = right_most_pntr[0];
+        input_t seq_right_most_value = right_most_pntr[seq_len - 1];
+
+        // Move src and dst pointers
+        src += block_offset + local_offset;
+        dst += block_offset + local_offset;
+
+        // Alpha and beta values for snake activatons. Applies exp by default
+        alpha = alpha + blockIdx.y;
+        input_t alpha_val = expf(alpha[0]);
+        beta = beta + blockIdx.y;
+        input_t beta_val = expf(beta[0]);
+
+        #pragma unroll
+        for (int it = 0; it < FILTER_SIZE; it += 1)
+        {
+            up_filter[it] = up_ftr[it];
+            down_filter[it] = down_ftr[it];
+        }
+
+        // Apply replication padding for upsampling, matching torch impl
+        #pragma unroll
+        for (int it = -HALF_FILTER_SIZE; it < BUFFER_SIZE + HALF_FILTER_SIZE; it += 1)
+        {
+            int element_index = seq_offset + it; // index for element
+            if ((element_index < 0) && (element_index >= -UPSAMPLE_REPLICATION_PAD))
+            {
+                elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_left_most_value;
+            }
+            if ((element_index >= seq_len) && (element_index < seq_len + UPSAMPLE_REPLICATION_PAD))
+            {
+                elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_right_most_value;
+            }
+            if ((element_index >= 0) && (element_index < seq_len))
+            {
+                elements[2 * (HALF_FILTER_SIZE + it)] = 2 * src[it];
+            }
+        }
+
+        // Apply upsampling strided convolution and write to intermediates. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT for replication padding of the downsampilng conv later
+        #pragma unroll
+        for (int it = 0; it < (2 * BUFFER_SIZE + 2 * FILTER_SIZE); it += 1)
+        {
+            input_t acc = 0.0;
+            int element_index = intermediate_seq_offset + it; // index for intermediate
+            #pragma unroll
+            for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1)
+            {
+                if ((element_index + f_idx) >= 0)
+                {
+                    acc += up_filter[f_idx] * elements[it + f_idx];
+                }
+            }
+            intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] = acc;
+        }
+
+        // Apply activation function. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT and DOWNSAMPLE_REPLICATION_PAD_RIGHT for replication padding of the downsampilng conv later
+        double no_div_by_zero = 0.000000001;
+        #pragma unroll
+        for (int it = 0; it < 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it += 1)
+        {
+            intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] += (1.0 / (beta_val + no_div_by_zero)) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val);
+        }
+
+        // Apply replication padding before downsampling conv from intermediates
+        #pragma unroll
+        for (int it = 0; it < DOWNSAMPLE_REPLICATION_PAD_LEFT; it += 1)
+        {
+            intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT];
+        }
+        #pragma unroll
+        for (int it = DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it < DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE + DOWNSAMPLE_REPLICATION_PAD_RIGHT; it += 1)
+        {
+            intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE - 1];
+        }
+
+        // Apply downsample strided convolution (assuming stride=2) from intermediates
+        #pragma unroll
+        for (int it = 0; it < BUFFER_SIZE; it += 1)
+        {
+            input_t acc = 0.0;
+            #pragma unroll
+            for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1)
+            {
+                // Add constant DOWNSAMPLE_REPLICATION_PAD_RIGHT to match torch implementation
+                acc += down_filter[f_idx] * intermediates[it * 2 + f_idx + DOWNSAMPLE_REPLICATION_PAD_RIGHT];
+            }
+            output[it] = acc;
+        }
+
+        // Write output to dst
+        #pragma unroll
+        for (int it = 0;  it < BUFFER_SIZE;  it += ELEMENTS_PER_LDG_STG)
+        {
+            int element_index = seq_offset + it;
+            if (element_index < seq_len)
+            {
+                dst[it] = output[it];
+            }
+        }
+
+    }
+
+    template <typename input_t, typename output_t, typename acc_t>
+    void dispatch_anti_alias_activation_forward(
+        output_t *dst,
+        const input_t *src,
+        const input_t *up_ftr,
+        const input_t *down_ftr,
+        const input_t *alpha,
+        const input_t *beta,
+        int batch_size,
+        int channels,
+        int seq_len)
+    {
+        if (seq_len == 0)
+        {
+            return;
+        }
+        else
+        {
+            // Use 128 threads per block to maximimize gpu utilization
+            constexpr int threads_per_block = 128;
+            constexpr int seq_len_per_block = 4096;
+            int blocks_per_seq_len = (seq_len + seq_len_per_block - 1) / seq_len_per_block;
+            dim3 blocks(blocks_per_seq_len, channels, batch_size);
+            dim3 threads(threads_per_block, 1, 1);
+
+            anti_alias_activation_forward<input_t, output_t, acc_t>
+                <<<blocks, threads, 0, at::cuda::getCurrentCUDAStream()>>>(dst, src, up_ftr, down_ftr, alpha, beta, batch_size, channels, seq_len);
+        }
+    }
+}
+
+extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta)
+{
+    // Input is a 3d tensor with dimensions [batches, channels, seq_len]
+    const int batches = input.size(0);
+    const int channels = input.size(1);
+    const int seq_len = input.size(2);
+
+    // Output
+    auto act_options = input.options().requires_grad(false);
+
+    torch::Tensor anti_alias_activation_results =
+        torch::empty({batches, channels, seq_len}, act_options);
+
+    void *input_ptr = static_cast<void *>(input.data_ptr());
+    void *up_filter_ptr = static_cast<void *>(up_filter.data_ptr());
+    void *down_filter_ptr = static_cast<void *>(down_filter.data_ptr());
+    void *alpha_ptr = static_cast<void *>(alpha.data_ptr());
+    void *beta_ptr = static_cast<void *>(beta.data_ptr());
+    void *anti_alias_activation_results_ptr = static_cast<void *>(anti_alias_activation_results.data_ptr());
+
+    DISPATCH_FLOAT_HALF_AND_BFLOAT(
+        input.scalar_type(),
+        "dispatch anti alias activation_forward",
+        dispatch_anti_alias_activation_forward<scalar_t, scalar_t, float>(
+            reinterpret_cast<scalar_t *>(anti_alias_activation_results_ptr),
+            reinterpret_cast<const scalar_t *>(input_ptr),
+            reinterpret_cast<const scalar_t *>(up_filter_ptr),
+            reinterpret_cast<const scalar_t *>(down_filter_ptr),
+            reinterpret_cast<const scalar_t *>(alpha_ptr),
+            reinterpret_cast<const scalar_t *>(beta_ptr),
+            batches,
+            channels,
+            seq_len););
+    return anti_alias_activation_results;
+}

selva_core/ext/bigvgan_v2/alias_free_activation/cuda/compat.h

@@ -0,0 +1,29 @@
+/* coding=utf-8
+ * Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/*This code is copied fron NVIDIA apex:
+ *     https://github.com/NVIDIA/apex
+ *     with minor changes. */
+
+#ifndef TORCH_CHECK
+#define TORCH_CHECK AT_CHECK
+#endif
+
+#ifdef VERSION_GE_1_3
+#define DATA_PTR data_ptr
+#else
+#define DATA_PTR data
+#endif

selva_core/ext/bigvgan_v2/alias_free_activation/cuda/load.py

@@ -0,0 +1,86 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+import os
+import pathlib
+import subprocess
+
+from torch.utils import cpp_extension
+
+"""
+Setting this param to a list has a problem of generating different compilation commands (with diferent order of architectures) and leading to recompilation of fused kernels. 
+Set it to empty stringo avoid recompilation and assign arch flags explicity in extra_cuda_cflags below
+"""
+os.environ["TORCH_CUDA_ARCH_LIST"] = ""
+
+
+def load():
+    # Check if cuda 11 is installed for compute capability 8.0
+    cc_flag = []
+    _, bare_metal_major, _ = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
+    if int(bare_metal_major) >= 11:
+        cc_flag.append("-gencode")
+        cc_flag.append("arch=compute_80,code=sm_80")
+
+    # Build path
+    srcpath = pathlib.Path(__file__).parent.absolute()
+    buildpath = srcpath / "build"
+    _create_build_dir(buildpath)
+
+    # Helper function to build the kernels.
+    def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
+        return cpp_extension.load(
+            name=name,
+            sources=sources,
+            build_directory=buildpath,
+            extra_cflags=[
+                "-O3",
+            ],
+            extra_cuda_cflags=[
+                "-O3",
+                "-gencode",
+                "arch=compute_70,code=sm_70",
+                "--use_fast_math",
+            ]
+            + extra_cuda_flags
+            + cc_flag,
+            verbose=True,
+        )
+
+    extra_cuda_flags = [
+        "-U__CUDA_NO_HALF_OPERATORS__",
+        "-U__CUDA_NO_HALF_CONVERSIONS__",
+        "--expt-relaxed-constexpr",
+        "--expt-extended-lambda",
+    ]
+
+    sources = [
+        srcpath / "anti_alias_activation.cpp",
+        srcpath / "anti_alias_activation_cuda.cu",
+    ]
+    anti_alias_activation_cuda = _cpp_extention_load_helper(
+        "anti_alias_activation_cuda", sources, extra_cuda_flags
+    )
+
+    return anti_alias_activation_cuda
+
+
+def _get_cuda_bare_metal_version(cuda_dir):
+    raw_output = subprocess.check_output(
+        [cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True
+    )
+    output = raw_output.split()
+    release_idx = output.index("release") + 1
+    release = output[release_idx].split(".")
+    bare_metal_major = release[0]
+    bare_metal_minor = release[1][0]
+
+    return raw_output, bare_metal_major, bare_metal_minor
+
+
+def _create_build_dir(buildpath):
+    try:
+        os.mkdir(buildpath)
+    except OSError:
+        if not os.path.isdir(buildpath):
+            print(f"Creation of the build directory {buildpath} failed")

selva_core/ext/bigvgan_v2/alias_free_activation/cuda/type_shim.h

@@ -0,0 +1,92 @@
+/* coding=utf-8
+ * Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <ATen/ATen.h>
+#include "compat.h"
+
+#define DISPATCH_FLOAT_HALF_AND_BFLOAT(TYPE, NAME, ...)                 \
+	switch (TYPE)                                                       \
+	{                                                                   \
+	case at::ScalarType::Float:                                         \
+	{                                                                   \
+		using scalar_t = float;                                         \
+		__VA_ARGS__;                                                    \
+		break;                                                          \
+	}                                                                   \
+	case at::ScalarType::Half:                                          \
+	{                                                                   \
+		using scalar_t = at::Half;                                      \
+		__VA_ARGS__;                                                    \
+		break;                                                          \
+	}                                                                   \
+	case at::ScalarType::BFloat16:                                      \
+	{                                                                   \
+		using scalar_t = at::BFloat16;                                  \
+		__VA_ARGS__;                                                    \
+		break;                                                          \
+	}                                                                   \
+	default:                                                            \
+		AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'"); \
+	}
+
+#define DISPATCH_FLOAT_HALF_AND_BFLOAT_INOUT_TYPES(TYPEIN, TYPEOUT, NAME, ...) \
+	switch (TYPEIN)                                                            \
+	{                                                                          \
+	case at::ScalarType::Float:                                                \
+	{                                                                          \
+		using scalar_t_in = float;                                             \
+		switch (TYPEOUT)                                                       \
+		{                                                                      \
+		case at::ScalarType::Float:                                            \
+		{                                                                      \
+			using scalar_t_out = float;                                        \
+			__VA_ARGS__;                                                       \
+			break;                                                             \
+		}                                                                      \
+		case at::ScalarType::Half:                                             \
+		{                                                                      \
+			using scalar_t_out = at::Half;                                     \
+			__VA_ARGS__;                                                       \
+			break;                                                             \
+		}                                                                      \
+		case at::ScalarType::BFloat16:                                         \
+		{                                                                      \
+			using scalar_t_out = at::BFloat16;                                 \
+			__VA_ARGS__;                                                       \
+			break;                                                             \
+		}                                                                      \
+		default:                                                               \
+			AT_ERROR(#NAME, " not implemented for '", toString(TYPEOUT), "'"); \
+		}                                                                      \
+		break;                                                                 \
+	}                                                                          \
+	case at::ScalarType::Half:                                                 \
+	{                                                                          \
+		using scalar_t_in = at::Half;                                          \
+		using scalar_t_out = at::Half;                                         \
+		__VA_ARGS__;                                                           \
+		break;                                                                 \
+	}                                                                          \
+	case at::ScalarType::BFloat16:                                             \
+	{                                                                          \
+		using scalar_t_in = at::BFloat16;                                      \
+		using scalar_t_out = at::BFloat16;                                     \
+		__VA_ARGS__;                                                           \
+		break;                                                                 \
+	}                                                                          \
+	default:                                                                   \
+		AT_ERROR(#NAME, " not implemented for '", toString(TYPEIN), "'");      \
+	}

selva_core/ext/bigvgan_v2/alias_free_activation/torch/__init__.py

@@ -0,0 +1,6 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+from .filter import *
+from .resample import *
+from .act import *

selva_core/ext/bigvgan_v2/alias_free_activation/torch/act.py

@@ -0,0 +1,32 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch.nn as nn
+
+from selva_core.ext.bigvgan_v2.alias_free_activation.torch.resample import (DownSample1d, UpSample1d)
+
+
+class Activation1d(nn.Module):
+
+    def __init__(
+        self,
+        activation,
+        up_ratio: int = 2,
+        down_ratio: int = 2,
+        up_kernel_size: int = 12,
+        down_kernel_size: int = 12,
+    ):
+        super().__init__()
+        self.up_ratio = up_ratio
+        self.down_ratio = down_ratio
+        self.act = activation
+        self.upsample = UpSample1d(up_ratio, up_kernel_size)
+        self.downsample = DownSample1d(down_ratio, down_kernel_size)
+
+    # x: [B,C,T]
+    def forward(self, x):
+        x = self.upsample(x)
+        x = self.act(x)
+        x = self.downsample(x)
+
+        return x

selva_core/ext/bigvgan_v2/alias_free_activation/torch/filter.py

@@ -0,0 +1,101 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+if "sinc" in dir(torch):
+    sinc = torch.sinc
+else:
+    # This code is adopted from adefossez's julius.core.sinc under the MIT License
+    # https://adefossez.github.io/julius/julius/core.html
+    #   LICENSE is in incl_licenses directory.
+    def sinc(x: torch.Tensor):
+        """
+        Implementation of sinc, i.e. sin(pi * x) / (pi * x)
+        __Warning__: Different to julius.sinc, the input is multiplied by `pi`!
+        """
+        return torch.where(
+            x == 0,
+            torch.tensor(1.0, device=x.device, dtype=x.dtype),
+            torch.sin(math.pi * x) / math.pi / x,
+        )
+
+
+# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License
+# https://adefossez.github.io/julius/julius/lowpass.html
+#   LICENSE is in incl_licenses directory.
+def kaiser_sinc_filter1d(
+    cutoff, half_width, kernel_size
+):  # return filter [1,1,kernel_size]
+    even = kernel_size % 2 == 0
+    half_size = kernel_size // 2
+
+    # For kaiser window
+    delta_f = 4 * half_width
+    A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95
+    if A > 50.0:
+        beta = 0.1102 * (A - 8.7)
+    elif A >= 21.0:
+        beta = 0.5842 * (A - 21) ** 0.4 + 0.07886 * (A - 21.0)
+    else:
+        beta = 0.0
+    window = torch.kaiser_window(kernel_size, beta=beta, periodic=False)
+
+    # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio
+    if even:
+        time = torch.arange(-half_size, half_size) + 0.5
+    else:
+        time = torch.arange(kernel_size) - half_size
+    if cutoff == 0:
+        filter_ = torch.zeros_like(time)
+    else:
+        filter_ = 2 * cutoff * window * sinc(2 * cutoff * time)
+        """
+        Normalize filter to have sum = 1, otherwise we will have a small leakage of the constant component in the input signal.
+        """
+        filter_ /= filter_.sum()
+        filter = filter_.view(1, 1, kernel_size)
+
+    return filter
+
+
+class LowPassFilter1d(nn.Module):
+    def __init__(
+        self,
+        cutoff=0.5,
+        half_width=0.6,
+        stride: int = 1,
+        padding: bool = True,
+        padding_mode: str = "replicate",
+        kernel_size: int = 12,
+    ):
+        """
+        kernel_size should be even number for stylegan3 setup, in this implementation, odd number is also possible.
+        """
+        super().__init__()
+        if cutoff < -0.0:
+            raise ValueError("Minimum cutoff must be larger than zero.")
+        if cutoff > 0.5:
+            raise ValueError("A cutoff above 0.5 does not make sense.")
+        self.kernel_size = kernel_size
+        self.even = kernel_size % 2 == 0
+        self.pad_left = kernel_size // 2 - int(self.even)
+        self.pad_right = kernel_size // 2
+        self.stride = stride
+        self.padding = padding
+        self.padding_mode = padding_mode
+        filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size)
+        self.register_buffer("filter", filter)
+
+    # Input [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        if self.padding:
+            x = F.pad(x, (self.pad_left, self.pad_right), mode=self.padding_mode)
+        out = F.conv1d(x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C)
+
+        return out

selva_core/ext/bigvgan_v2/alias_free_activation/torch/resample.py

@@ -0,0 +1,54 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch.nn as nn
+from torch.nn import functional as F
+
+from selva_core.ext.bigvgan_v2.alias_free_activation.torch.filter import (LowPassFilter1d,
+                                                                       kaiser_sinc_filter1d)
+
+
+class UpSample1d(nn.Module):
+
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = (int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size)
+        self.stride = ratio
+        self.pad = self.kernel_size // ratio - 1
+        self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2
+        self.pad_right = (self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2)
+        filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio,
+                                      half_width=0.6 / ratio,
+                                      kernel_size=self.kernel_size)
+        self.register_buffer("filter", filter)
+
+    # x: [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        x = F.pad(x, (self.pad, self.pad), mode="replicate")
+        x = self.ratio * F.conv_transpose1d(
+            x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C)
+        x = x[..., self.pad_left:-self.pad_right]
+
+        return x
+
+
+class DownSample1d(nn.Module):
+
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = (int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size)
+        self.lowpass = LowPassFilter1d(
+            cutoff=0.5 / ratio,
+            half_width=0.6 / ratio,
+            stride=ratio,
+            kernel_size=self.kernel_size,
+        )
+
+    def forward(self, x):
+        xx = self.lowpass(x)
+
+        return xx

selva_core/ext/bigvgan_v2/bigvgan.py

@@ -0,0 +1,439 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import json
+import os
+from pathlib import Path
+from typing import Dict, Optional, Union
+
+import torch
+import torch.nn as nn
+from huggingface_hub import PyTorchModelHubMixin, hf_hub_download
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils.parametrizations import weight_norm
+from torch.nn.utils.parametrize import remove_parametrizations
+
+from selva_core.ext.bigvgan_v2 import activations
+from selva_core.ext.bigvgan_v2.alias_free_activation.torch.act import \
+    Activation1d as TorchActivation1d
+from selva_core.ext.bigvgan_v2.env import AttrDict
+from selva_core.ext.bigvgan_v2.utils import get_padding, init_weights
+
+
+def load_hparams_from_json(path) -> AttrDict:
+    with open(path) as f:
+        data = f.read()
+    return AttrDict(json.loads(data))
+
+
+class AMPBlock1(torch.nn.Module):
+    """
+    AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer.
+    AMPBlock1 has additional self.convs2 that contains additional Conv1d layers with a fixed dilation=1 followed by each layer in self.convs1
+
+    Args:
+        h (AttrDict): Hyperparameters.
+        channels (int): Number of convolution channels.
+        kernel_size (int): Size of the convolution kernel. Default is 3.
+        dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5).
+        activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None.
+    """
+
+    def __init__(
+            self,
+            h: AttrDict,
+            channels: int,
+            kernel_size: int = 3,
+            dilation: tuple = (1, 3, 5),
+            activation: str = None,
+    ):
+        super().__init__()
+
+        self.h = h
+
+        self.convs1 = nn.ModuleList([
+            weight_norm(
+                Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    stride=1,
+                    dilation=d,
+                    padding=get_padding(kernel_size, d),
+                )) for d in dilation
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList([
+            weight_norm(
+                Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    stride=1,
+                    dilation=1,
+                    padding=get_padding(kernel_size, 1),
+                )) for _ in range(len(dilation))
+        ])
+        self.convs2.apply(init_weights)
+
+        self.num_layers = len(self.convs1) + len(self.convs2)  # Total number of conv layers
+
+        # Select which Activation1d, lazy-load cuda version to ensure backward compatibility
+        if self.h.get("use_cuda_kernel", False):
+            from alias_free_activation.cuda.activation1d import \
+                Activation1d as CudaActivation1d
+
+            Activation1d = CudaActivation1d
+        else:
+            Activation1d = TorchActivation1d
+
+        # Activation functions
+        if activation == "snake":
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == "snakebeta":
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        acts1, acts2 = self.activations[::2], self.activations[1::2]
+        for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2):
+            xt = a1(x)
+            xt = c1(xt)
+            xt = a2(xt)
+            xt = c2(xt)
+            x = xt + x
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_parametrizations(l, 'weight')
+        for l in self.convs2:
+            remove_parametrizations(l, 'weight')
+
+
+class AMPBlock2(torch.nn.Module):
+    """
+    AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer.
+    Unlike AMPBlock1, AMPBlock2 does not contain extra Conv1d layers with fixed dilation=1
+
+    Args:
+        h (AttrDict): Hyperparameters.
+        channels (int): Number of convolution channels.
+        kernel_size (int): Size of the convolution kernel. Default is 3.
+        dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5).
+        activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None.
+    """
+
+    def __init__(
+            self,
+            h: AttrDict,
+            channels: int,
+            kernel_size: int = 3,
+            dilation: tuple = (1, 3, 5),
+            activation: str = None,
+    ):
+        super().__init__()
+
+        self.h = h
+
+        self.convs = nn.ModuleList([
+            weight_norm(
+                Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    stride=1,
+                    dilation=d,
+                    padding=get_padding(kernel_size, d),
+                )) for d in dilation
+        ])
+        self.convs.apply(init_weights)
+
+        self.num_layers = len(self.convs)  # Total number of conv layers
+
+        # Select which Activation1d, lazy-load cuda version to ensure backward compatibility
+        if self.h.get("use_cuda_kernel", False):
+            from alias_free_activation.cuda.activation1d import \
+                Activation1d as CudaActivation1d
+
+            Activation1d = CudaActivation1d
+        else:
+            Activation1d = TorchActivation1d
+
+        # Activation functions
+        if activation == "snake":
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == "snakebeta":
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        for c, a in zip(self.convs, self.activations):
+            xt = a(x)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class BigVGAN(
+        torch.nn.Module,
+        PyTorchModelHubMixin,
+        library_name="bigvgan",
+        repo_url="https://github.com/NVIDIA/BigVGAN",
+        docs_url="https://github.com/NVIDIA/BigVGAN/blob/main/README.md",
+        pipeline_tag="audio-to-audio",
+        license="mit",
+        tags=["neural-vocoder", "audio-generation", "arxiv:2206.04658"],
+):
+    """
+    BigVGAN is a neural vocoder model that applies anti-aliased periodic activation for residual blocks (resblocks).
+    New in BigVGAN-v2: it can optionally use optimized CUDA kernels for AMP (anti-aliased multi-periodicity) blocks.
+
+    Args:
+        h (AttrDict): Hyperparameters.
+        use_cuda_kernel (bool): If set to True, loads optimized CUDA kernels for AMP. This should be used for inference only, as training is not supported with CUDA kernels.
+
+    Note:
+        - The `use_cuda_kernel` parameter should be used for inference only, as training with CUDA kernels is not supported.
+        - Ensure that the activation function is correctly specified in the hyperparameters (h.activation).
+    """
+
+    def __init__(self, h: AttrDict, use_cuda_kernel: bool = False):
+        super().__init__()
+        self.h = h
+        self.h["use_cuda_kernel"] = use_cuda_kernel
+
+        # Select which Activation1d, lazy-load cuda version to ensure backward compatibility
+        if self.h.get("use_cuda_kernel", False):
+            from alias_free_activation.cuda.activation1d import \
+                Activation1d as CudaActivation1d
+
+            Activation1d = CudaActivation1d
+        else:
+            Activation1d = TorchActivation1d
+
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+
+        # Pre-conv
+        self.conv_pre = weight_norm(Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3))
+
+        # Define which AMPBlock to use. BigVGAN uses AMPBlock1 as default
+        if h.resblock == "1":
+            resblock_class = AMPBlock1
+        elif h.resblock == "2":
+            resblock_class = AMPBlock2
+        else:
+            raise ValueError(
+                f"Incorrect resblock class specified in hyperparameters. Got {h.resblock}")
+
+        # Transposed conv-based upsamplers. does not apply anti-aliasing
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                nn.ModuleList([
+                    weight_norm(
+                        ConvTranspose1d(
+                            h.upsample_initial_channel // (2**i),
+                            h.upsample_initial_channel // (2**(i + 1)),
+                            k,
+                            u,
+                            padding=(k - u) // 2,
+                        ))
+                ]))
+
+        # Residual blocks using anti-aliased multi-periodicity composition modules (AMP)
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2**(i + 1))
+            for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
+                self.resblocks.append(resblock_class(h, ch, k, d, activation=h.activation))
+
+        # Post-conv
+        activation_post = (activations.Snake(ch, alpha_logscale=h.snake_logscale)
+                           if h.activation == "snake" else
+                           (activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale)
+                            if h.activation == "snakebeta" else None))
+        if activation_post is None:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+        self.activation_post = Activation1d(activation=activation_post)
+
+        # Whether to use bias for the final conv_post. Default to True for backward compatibility
+        self.use_bias_at_final = h.get("use_bias_at_final", True)
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3, bias=self.use_bias_at_final))
+
+        # Weight initialization
+        for i in range(len(self.ups)):
+            self.ups[i].apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+        # Final tanh activation. Defaults to True for backward compatibility
+        self.use_tanh_at_final = h.get("use_tanh_at_final", True)
+
+    def forward(self, x):
+        # Pre-conv
+        x = self.conv_pre(x)
+
+        for i in range(self.num_upsamples):
+            # Upsampling
+            for i_up in range(len(self.ups[i])):
+                x = self.ups[i][i_up](x)
+            # AMP blocks
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        # Post-conv
+        x = self.activation_post(x)
+        x = self.conv_post(x)
+        # Final tanh activation
+        if self.use_tanh_at_final:
+            x = torch.tanh(x)
+        else:
+            x = torch.clamp(x, min=-1.0, max=1.0)  # Bound the output to [-1, 1]
+
+        return x
+
+    def remove_weight_norm(self):
+        try:
+            print("Removing weight norm...")
+            for l in self.ups:
+                for l_i in l:
+                    remove_parametrizations(l_i, 'weight')
+            for l in self.resblocks:
+                l.remove_weight_norm()
+            remove_parametrizations(self.conv_pre, 'weight')
+            remove_parametrizations(self.conv_post, 'weight')
+        except ValueError:
+            print("[INFO] Model already removed weight norm. Skipping!")
+            pass
+
+    # Additional methods for huggingface_hub support
+    def _save_pretrained(self, save_directory: Path) -> None:
+        """Save weights and config.json from a Pytorch model to a local directory."""
+
+        model_path = save_directory / "bigvgan_generator.pt"
+        torch.save({"generator": self.state_dict()}, model_path)
+
+        config_path = save_directory / "config.json"
+        with open(config_path, "w") as config_file:
+            json.dump(self.h, config_file, indent=4)
+
+    @classmethod
+    def _from_pretrained(
+        cls,
+        *,
+        model_id: str,
+        revision: str,
+        cache_dir: str,
+        force_download: bool,
+        proxies: Optional[Dict],
+        resume_download: bool,
+        local_files_only: bool,
+        token: Union[str, bool, None],
+        map_location: str = "cpu",  # Additional argument
+        strict: bool = False,  # Additional argument
+        use_cuda_kernel: bool = False,
+        **model_kwargs,
+    ):
+        """Load Pytorch pretrained weights and return the loaded model."""
+
+        # Download and load hyperparameters (h) used by BigVGAN
+        if os.path.isdir(model_id):
+            print("Loading config.json from local directory")
+            config_file = os.path.join(model_id, "config.json")
+        else:
+            config_file = hf_hub_download(
+                repo_id=model_id,
+                filename="config.json",
+                revision=revision,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                token=token,
+                local_files_only=local_files_only,
+            )
+        h = load_hparams_from_json(config_file)
+
+        # instantiate BigVGAN using h
+        if use_cuda_kernel:
+            print(
+                f"[WARNING] You have specified use_cuda_kernel=True during BigVGAN.from_pretrained(). Only inference is supported (training is not implemented)!"
+            )
+            print(
+                f"[WARNING] You need nvcc and ninja installed in your system that matches your PyTorch build is using to build the kernel. If not, the model will fail to initialize or generate incorrect waveform!"
+            )
+            print(
+                f"[WARNING] For detail, see the official GitHub repository: https://github.com/NVIDIA/BigVGAN?tab=readme-ov-file#using-custom-cuda-kernel-for-synthesis"
+            )
+        model = cls(h, use_cuda_kernel=use_cuda_kernel)
+
+        # Download and load pretrained generator weight
+        if os.path.isdir(model_id):
+            print("Loading weights from local directory")
+            model_file = os.path.join(model_id, "bigvgan_generator.pt")
+        else:
+            print(f"Loading weights from {model_id}")
+            model_file = hf_hub_download(
+                repo_id=model_id,
+                filename="bigvgan_generator.pt",
+                revision=revision,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                token=token,
+                local_files_only=local_files_only,
+            )
+
+        checkpoint_dict = torch.load(model_file, map_location=map_location, weights_only=True)
+
+        try:
+            model.load_state_dict(checkpoint_dict["generator"])
+        except RuntimeError:
+            print(
+                f"[INFO] the pretrained checkpoint does not contain weight norm. Loading the checkpoint after removing weight norm!"
+            )
+            model.remove_weight_norm()
+            model.load_state_dict(checkpoint_dict["generator"])
+
+        return model

selva_core/ext/bigvgan_v2/env.py

@@ -0,0 +1,18 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import os
+import shutil
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+def build_env(config, config_name, path):
+    t_path = os.path.join(path, config_name)
+    if config != t_path:
+        os.makedirs(path, exist_ok=True)
+        shutil.copyfile(config, os.path.join(path, config_name))

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_1

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Jungil Kong
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_2

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Edward Dixon
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_3

@@ -0,0 +1,201 @@
+                                Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_4

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2019, Seungwon Park 박승원
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_5

@@ -0,0 +1,16 @@
+Copyright 2020 Alexandre Défossez
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
+associated documentation files (the "Software"), to deal in the Software without restriction,
+including without limitation the rights to use, copy, modify, merge, publish, distribute,
+sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or
+substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT
+NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
+DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_6

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023-present, Descript
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_7

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Charactr Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan_v2/incl_licenses/LICENSE_8

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Amphion
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/bigvgan_v2/utils.py

@@ -0,0 +1,31 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import os
+
+import torch
+from torch.nn.utils import weight_norm
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def apply_weight_norm(m):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        weight_norm(m)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print(f"Loading '{filepath}'")
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict

selva_core/ext/mel_converter.py

@@ -0,0 +1,139 @@
+# Reference: # https://github.com/bytedance/Make-An-Audio-2
+from typing import Literal
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+def librosa_mel_fn(*, sr, n_fft, n_mels=128, fmin=0.0, fmax=None):
+    """Pure-numpy mel filterbank equivalent to librosa.filters.mel (HTK scale, no norm).
+
+    Replaces the librosa import to avoid the librosa → numba → NumPy-version
+    incompatibility that exists in some ComfyUI environments.
+    """
+    if fmax is None:
+        fmax = sr / 2.0
+
+    def hz_to_mel(f):
+        return 2595.0 * np.log10(1.0 + np.asarray(f) / 700.0)
+
+    def mel_to_hz(m):
+        return 700.0 * (10.0 ** (np.asarray(m) / 2595.0) - 1.0)
+
+    n_freqs = n_fft // 2 + 1
+    fft_freqs = np.linspace(0.0, sr / 2.0, n_freqs)
+
+    mel_min = hz_to_mel(fmin)
+    mel_max = hz_to_mel(fmax)
+    mel_points = np.linspace(mel_min, mel_max, n_mels + 2)
+    hz_points = mel_to_hz(mel_points)
+
+    weights = np.zeros((n_mels, n_freqs), dtype=np.float32)
+    for m in range(1, n_mels + 1):
+        f_lo, f_mid, f_hi = hz_points[m - 1], hz_points[m], hz_points[m + 1]
+        up   = (fft_freqs - f_lo) / (f_mid - f_lo + 1e-12)
+        down = (f_hi - fft_freqs) / (f_hi - f_mid + 1e-12)
+        weights[m - 1] = np.maximum(0.0, np.minimum(up, down))
+
+    return weights
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, *, norm_fn):
+    return norm_fn(torch.clamp(x, min=clip_val) * C)
+
+
+def spectral_normalize_torch(magnitudes, norm_fn):
+    output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn)
+    return output
+
+
+class MelConverter(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        sampling_rate: float,
+        n_fft: int,
+        num_mels: int,
+        hop_size: int,
+        win_size: int,
+        fmin: float,
+        fmax: float,
+        norm_fn,
+    ):
+        super().__init__()
+        self.sampling_rate = sampling_rate
+        self.n_fft = n_fft
+        self.num_mels = num_mels
+        self.hop_size = hop_size
+        self.win_size = win_size
+        self.fmin = fmin
+        self.fmax = fmax
+        self.norm_fn = norm_fn
+
+        mel = librosa_mel_fn(sr=self.sampling_rate,
+                             n_fft=self.n_fft,
+                             n_mels=self.num_mels,
+                             fmin=self.fmin,
+                             fmax=self.fmax)
+        mel_basis = torch.from_numpy(mel).float()
+        hann_window = torch.hann_window(self.win_size)
+
+        self.register_buffer('mel_basis', mel_basis)
+        self.register_buffer('hann_window', hann_window)
+
+    @property
+    def device(self):
+        return self.mel_basis.device
+
+    def forward(self, waveform: torch.Tensor, center: bool = False) -> torch.Tensor:
+        waveform = waveform.clamp(min=-1., max=1.).to(self.device)
+
+        waveform = torch.nn.functional.pad(
+            waveform.unsqueeze(1),
+            [int((self.n_fft - self.hop_size) / 2),
+             int((self.n_fft - self.hop_size) / 2)],
+            mode='reflect')
+        waveform = waveform.squeeze(1)
+
+        spec = torch.stft(waveform,
+                          self.n_fft,
+                          hop_length=self.hop_size,
+                          win_length=self.win_size,
+                          window=self.hann_window,
+                          center=center,
+                          pad_mode='reflect',
+                          normalized=False,
+                          onesided=True,
+                          return_complex=True)
+
+        spec = torch.view_as_real(spec)
+        spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+        spec = torch.matmul(self.mel_basis, spec)
+        spec = spectral_normalize_torch(spec, self.norm_fn)
+
+        return spec
+
+
+def get_mel_converter(mode: Literal['16k', '44k']) -> MelConverter:
+    if mode == '16k':
+        return MelConverter(sampling_rate=16_000,
+                            n_fft=1024,
+                            num_mels=80,
+                            hop_size=256,
+                            win_size=1024,
+                            fmin=0,
+                            fmax=8_000,
+                            norm_fn=torch.log10)
+    elif mode == '44k':
+        return MelConverter(sampling_rate=44_100,
+                            n_fft=2048,
+                            num_mels=128,
+                            hop_size=512,
+                            win_size=2048,
+                            fmin=0,
+                            fmax=44100 / 2,
+                            norm_fn=torch.log)
+    else:
+        raise ValueError(f'Unknown mode: {mode}')

selva_core/ext/rotary_embeddings.py

@@ -0,0 +1,35 @@
+from typing import Union
+
+import torch
+from einops import rearrange
+from torch import Tensor
+
+# Ref: https://github.com/black-forest-labs/flux/blob/main/src/flux/math.py
+# Ref: https://github.com/lucidrains/rotary-embedding-torch
+
+
+def compute_rope_rotations(length: int,
+                           dim: int,
+                           theta: int,
+                           *,
+                           freq_scaling: float = 1.0,
+                           device: Union[torch.device, str] = 'cpu') -> Tensor:
+    assert dim % 2 == 0
+
+    with torch.amp.autocast(device_type='cuda', enabled=False):
+        pos = torch.arange(length, dtype=torch.float32, device=device)
+        freqs = 1.0 / (theta**(torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
+        freqs *= freq_scaling
+
+        rot = torch.einsum('..., f -> ... f', pos, freqs)
+        rot = torch.stack([torch.cos(rot), -torch.sin(rot), torch.sin(rot), torch.cos(rot)], dim=-1)
+        rot = rearrange(rot, 'n d (i j) -> 1 n d i j', i=2, j=2)
+        return rot
+
+
+def apply_rope(x: Tensor, rot: Tensor) -> tuple[Tensor, Tensor]:
+    with torch.amp.autocast(device_type='cuda', enabled=False):
+        _x = x.float()
+        _x = _x.view(*_x.shape[:-1], -1, 1, 2)
+        x_out = rot[..., 0] * _x[..., 0] + rot[..., 1] * _x[..., 1]
+        return x_out.reshape(*x.shape).to(dtype=x.dtype)

selva_core/ext/stft_converter.py

@@ -0,0 +1,183 @@
+# Reference: # https://github.com/bytedance/Make-An-Audio-2
+
+import torch
+import torch.nn as nn
+import torchaudio
+from einops import rearrange
+from librosa.filters import mel as librosa_mel_fn
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, norm_fn=torch.log10):
+    return norm_fn(torch.clamp(x, min=clip_val) * C)
+
+
+def spectral_normalize_torch(magnitudes, norm_fn):
+    output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn)
+    return output
+
+
+class STFTConverter(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        sampling_rate: float = 16_000,
+        n_fft: int = 1024,
+        num_mels: int = 128,
+        hop_size: int = 256,
+        win_size: int = 1024,
+        fmin: float = 0,
+        fmax: float = 8_000,
+        norm_fn=torch.log,
+    ):
+        super().__init__()
+        self.sampling_rate = sampling_rate
+        self.n_fft = n_fft
+        self.num_mels = num_mels
+        self.hop_size = hop_size
+        self.win_size = win_size
+        self.fmin = fmin
+        self.fmax = fmax
+        self.norm_fn = norm_fn
+
+        mel = librosa_mel_fn(sr=self.sampling_rate,
+                             n_fft=self.n_fft,
+                             n_mels=self.num_mels,
+                             fmin=self.fmin,
+                             fmax=self.fmax)
+        mel_basis = torch.from_numpy(mel).float()
+        hann_window = torch.hann_window(self.win_size)
+
+        self.register_buffer('mel_basis', mel_basis)
+        self.register_buffer('hann_window', hann_window)
+
+    @property
+    def device(self):
+        return self.hann_window.device
+
+    def forward(self, waveform: torch.Tensor) -> torch.Tensor:
+        # input: batch_size * length
+        bs = waveform.shape[0]
+        waveform = waveform.clamp(min=-1., max=1.)
+
+        spec = torch.stft(waveform,
+                          self.n_fft,
+                          hop_length=self.hop_size,
+                          win_length=self.win_size,
+                          window=self.hann_window,
+                          center=True,
+                          pad_mode='reflect',
+                          normalized=False,
+                          onesided=True,
+                          return_complex=True)
+
+        spec = torch.view_as_real(spec)
+        # print('After stft', spec.shape, spec.min(), spec.max(), spec.mean())
+
+        power = spec.pow(2).sum(-1)
+        angle = torch.atan2(spec[..., 1], spec[..., 0])
+
+        print('power', power.shape, power.min(), power.max(), power.mean())
+        print('angle', angle.shape, angle.min(), angle.max(), angle.mean())
+
+        # print('mel', self.mel_basis.shape, self.mel_basis.min(), self.mel_basis.max(),
+        #       self.mel_basis.mean())
+
+        # spec = rearrange(spec, 'b f t c -> (b c) f t')
+
+        # spec = self.mel_transform(spec)
+
+        # spec = torch.matmul(self.mel_basis, spec)
+
+        # print('After mel', spec.shape, spec.min(), spec.max(), spec.mean())
+
+        # spec = spectral_normalize_torch(spec, self.norm_fn)
+
+        # print('After norm', spec.shape, spec.min(), spec.max(), spec.mean())
+
+        # compute magnitude
+        # magnitude = torch.sqrt((spec**2).sum(-1))
+        # normalize by magnitude
+        # scaled_magnitude = torch.log10(magnitude.clamp(min=1e-5)) * 10
+        # spec = spec / magnitude.unsqueeze(-1) * scaled_magnitude.unsqueeze(-1)
+
+        # power = torch.log10(power.clamp(min=1e-5)) * 10
+        power = torch.log10(power.clamp(min=1e-5))
+
+        print('After scaling', power.shape, power.min(), power.max(), power.mean())
+
+        spec = torch.stack([power, angle], dim=-1)
+
+        # spec = rearrange(spec, '(b c) f t -> b c f t', b=bs)
+        spec = rearrange(spec, 'b f t c -> b c f t', b=bs)
+
+        # spec[:, :, 400:] = 0
+
+        return spec
+
+    def invert(self, spec: torch.Tensor, length: int) -> torch.Tensor:
+        bs = spec.shape[0]
+
+        # spec = rearrange(spec, 'b c f t -> (b c) f t')
+        # print(spec.shape, self.mel_basis.shape)
+        # spec = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), spec).solution
+        # spec = torch.linalg.pinv(self.mel_basis.unsqueeze(0)) @ spec
+
+        # spec = self.invmel_transform(spec)
+
+        spec = rearrange(spec, 'b c f t -> b f t c', b=bs).contiguous()
+
+        # spec[..., 0] = 10**(spec[..., 0] / 10)
+
+        power = spec[..., 0]
+        power = 10**power
+
+        # print('After unscaling', spec[..., 0].shape, spec[..., 0].min(), spec[..., 0].max(),
+        #       spec[..., 0].mean())
+
+        unit_vector = torch.stack([
+            torch.cos(spec[..., 1]),
+            torch.sin(spec[..., 1]),
+        ], dim=-1)
+
+        spec = torch.sqrt(power) * unit_vector
+
+        # spec = rearrange(spec, '(b c) f t -> b f t c', b=bs).contiguous()
+        spec = torch.view_as_complex(spec)
+
+        waveform = torch.istft(
+            spec,
+            self.n_fft,
+            length=length,
+            hop_length=self.hop_size,
+            win_length=self.win_size,
+            window=self.hann_window,
+            center=True,
+            normalized=False,
+            onesided=True,
+            return_complex=False,
+        )
+
+        return waveform
+
+
+if __name__ == '__main__':
+
+    converter = STFTConverter(sampling_rate=16000)
+
+    signal = torchaudio.load('./output/ZZ6GRocWW38_000090.wav')[0]
+    # resample signal at 44100 Hz
+    # signal = torchaudio.transforms.Resample(16_000, 44_100)(signal)
+
+    L = signal.shape[1]
+    print('Input signal', signal.shape)
+    spec = converter(signal)
+
+    print('Final spec', spec.shape)
+
+    signal_recon = converter.invert(spec, length=L)
+    print('Output signal', signal_recon.shape, signal_recon.min(), signal_recon.max(),
+          signal_recon.mean())
+
+    print('MSE', torch.nn.functional.mse_loss(signal, signal_recon))
+    torchaudio.save('./output/ZZ6GRocWW38_000090_recon.wav', signal_recon, 16000)

selva_core/ext/stft_converter_mel.py

@@ -0,0 +1,234 @@
+# Reference: # https://github.com/bytedance/Make-An-Audio-2
+
+import torch
+import torch.nn as nn
+import torchaudio
+from einops import rearrange
+from librosa.filters import mel as librosa_mel_fn
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5, norm_fn=torch.log10):
+    return norm_fn(torch.clamp(x, min=clip_val) * C)
+
+
+def spectral_normalize_torch(magnitudes, norm_fn):
+    output = dynamic_range_compression_torch(magnitudes, norm_fn=norm_fn)
+    return output
+
+
+class STFTConverter(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        sampling_rate: float = 16_000,
+        n_fft: int = 1024,
+        num_mels: int = 128,
+        hop_size: int = 256,
+        win_size: int = 1024,
+        fmin: float = 0,
+        fmax: float = 8_000,
+        norm_fn=torch.log,
+    ):
+        super().__init__()
+        self.sampling_rate = sampling_rate
+        self.n_fft = n_fft
+        self.num_mels = num_mels
+        self.hop_size = hop_size
+        self.win_size = win_size
+        self.fmin = fmin
+        self.fmax = fmax
+        self.norm_fn = norm_fn
+
+        mel = librosa_mel_fn(sr=self.sampling_rate,
+                             n_fft=self.n_fft,
+                             n_mels=self.num_mels,
+                             fmin=self.fmin,
+                             fmax=self.fmax)
+        mel_basis = torch.from_numpy(mel).float()
+        hann_window = torch.hann_window(self.win_size)
+
+        self.register_buffer('mel_basis', mel_basis)
+        self.register_buffer('hann_window', hann_window)
+
+    @property
+    def device(self):
+        return self.hann_window.device
+
+    def forward(self, waveform: torch.Tensor) -> torch.Tensor:
+        # input: batch_size * length
+        bs = waveform.shape[0]
+        waveform = waveform.clamp(min=-1., max=1.)
+
+        spec = torch.stft(waveform,
+                          self.n_fft,
+                          hop_length=self.hop_size,
+                          win_length=self.win_size,
+                          window=self.hann_window,
+                          center=True,
+                          pad_mode='reflect',
+                          normalized=False,
+                          onesided=True,
+                          return_complex=True)
+
+        spec = torch.view_as_real(spec)
+        # print('After stft', spec.shape, spec.min(), spec.max(), spec.mean())
+
+        power = (spec.pow(2).sum(-1))**(0.5)
+        angle = torch.atan2(spec[..., 1], spec[..., 0])
+
+        print('power 1', power.shape, power.min(), power.max(), power.mean())
+        print('angle 1', angle.shape, angle.min(), angle.max(), angle.mean(), angle[:, :2, :2])
+
+        # print('mel', self.mel_basis.shape, self.mel_basis.min(), self.mel_basis.max(),
+        #       self.mel_basis.mean())
+
+        # spec = self.mel_transform(spec)
+
+        # power = torch.matmul(self.mel_basis, power)
+
+        spec = rearrange(spec, 'b f t c -> (b c) f t')
+        spec = self.mel_basis.unsqueeze(0) @ spec
+        spec = rearrange(spec, '(b c) f t -> b f t c', b=bs)
+
+        power = (spec.pow(2).sum(-1))**(0.5)
+        angle = torch.atan2(spec[..., 1], spec[..., 0])
+
+        print('power', power.shape, power.min(), power.max(), power.mean())
+        print('angle', angle.shape, angle.min(), angle.max(), angle.mean(), angle[:, :2, :2])
+
+        # print('After mel', spec.shape, spec.min(), spec.max(), spec.mean())
+
+        # spec = spectral_normalize_torch(spec, self.norm_fn)
+
+        # print('After norm', spec.shape, spec.min(), spec.max(), spec.mean())
+
+        # compute magnitude
+        # magnitude = torch.sqrt((spec**2).sum(-1))
+        # normalize by magnitude
+        # scaled_magnitude = torch.log10(magnitude.clamp(min=1e-5)) * 10
+        # spec = spec / magnitude.unsqueeze(-1) * scaled_magnitude.unsqueeze(-1)
+
+        # power = torch.log10(power.clamp(min=1e-5)) * 10
+        power = torch.log10(power.clamp(min=1e-8))
+
+        print('After scaling', power.shape, power.min(), power.max(), power.mean())
+
+        # spec = torch.stack([power, angle], dim=-1)
+
+        # spec = rearrange(spec, '(b c) f t -> b c f t', b=bs)
+        # spec = rearrange(spec, 'b f t c -> b c f t', b=bs)
+
+        # spec[:, :, 400:] = 0
+
+        return power, angle
+        # return spec[..., 0], spec[..., 1]
+
+    def invert(self, spec: torch.Tensor, length: int) -> torch.Tensor:
+
+        power, angle = spec
+
+        bs = power.shape[0]
+
+        # spec = rearrange(spec, 'b c f t -> (b c) f t')
+        # print(spec.shape, self.mel_basis.shape)
+        # spec = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), spec).solution
+        # spec = torch.linalg.pinv(self.mel_basis.unsqueeze(0)) @ spec
+
+        # spec = self.invmel_transform(spec)
+
+        # spec = rearrange(spec, 'b c f t -> b f t c', b=bs).contiguous()
+
+        # spec[..., 0] = 10**(spec[..., 0] / 10)
+
+        # power = spec[..., 0]
+        power = 10**power
+
+        # print('After unscaling', spec[..., 0].shape, spec[..., 0].min(), spec[..., 0].max(),
+        #       spec[..., 0].mean())
+
+        unit_vector = torch.stack([
+            torch.cos(angle),
+            torch.sin(angle),
+        ], dim=-1)
+
+        spec = power.unsqueeze(-1) * unit_vector
+
+        # power = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), power).solution
+        spec = rearrange(spec, 'b f t c -> (b c) f t')
+        spec = torch.linalg.pinv(self.mel_basis.unsqueeze(0)) @ spec
+        # spec = torch.linalg.lstsq(self.mel_basis.unsqueeze(0), spec).solution
+        spec = rearrange(spec, '(b c) f t -> b f t c', b=bs).contiguous()
+
+        power = (spec.pow(2).sum(-1))**(0.5)
+        angle = torch.atan2(spec[..., 1], spec[..., 0])
+
+        print('power 2', power.shape, power.min(), power.max(), power.mean())
+        print('angle 2', angle.shape, angle.min(), angle.max(), angle.mean(), angle[:, :2, :2])
+
+        # spec = rearrange(spec, '(b c) f t -> b f t c', b=bs).contiguous()
+        spec = torch.view_as_complex(spec)
+
+        waveform = torch.istft(
+            spec,
+            self.n_fft,
+            length=length,
+            hop_length=self.hop_size,
+            win_length=self.win_size,
+            window=self.hann_window,
+            center=True,
+            normalized=False,
+            onesided=True,
+            return_complex=False,
+        )
+
+        return waveform
+
+
+if __name__ == '__main__':
+
+    converter = STFTConverter(sampling_rate=16000)
+
+    signal = torchaudio.load('./output/ZZ6GRocWW38_000090.wav')[0]
+    # resample signal at 44100 Hz
+    # signal = torchaudio.transforms.Resample(16_000, 44_100)(signal)
+
+    L = signal.shape[1]
+    print('Input signal', signal.shape)
+    spec = converter(signal)
+
+    power, angle = spec
+
+    # print(power.shape, angle.shape)
+    # print(power, power.min(), power.max(), power.mean())
+    # power = power.clamp(-1, 1)
+    # angle = angle.clamp(-1, 1)
+
+    import matplotlib.pyplot as plt
+
+    # Visualize power
+    plt.figure()
+    plt.imshow(power[0].detach().numpy(), aspect='auto', origin='lower')
+    plt.colorbar()
+    plt.title('Power')
+    plt.xlabel('Time')
+    plt.ylabel('Frequency')
+    plt.savefig('./output/power.png')
+
+    # Visualize angle
+    plt.figure()
+    plt.imshow(angle[0].detach().numpy(), aspect='auto', origin='lower')
+    plt.colorbar()
+    plt.title('Angle')
+    plt.xlabel('Time')
+    plt.ylabel('Frequency')
+    plt.savefig('./output/angle.png')
+
+    # print('Final spec', spec.shape)
+
+    signal_recon = converter.invert(spec, length=L)
+    print('Output signal', signal_recon.shape, signal_recon.min(), signal_recon.max(),
+          signal_recon.mean())
+
+    print('MSE', torch.nn.functional.mse_loss(signal, signal_recon))
+    torchaudio.save('./output/ZZ6GRocWW38_000090_recon.wav', signal_recon, 16000)

selva_core/ext/synchformer/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Vladimir Iashin
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

selva_core/ext/synchformer/__init__.py

@@ -0,0 +1 @@
+from selva_core.ext.synchformer.synchformer import Synchformer

selva_core/ext/synchformer/astransformer.py

@@ -0,0 +1,279 @@
+import logging
+import torch
+from torch import nn
+# importing modified version of AST
+from transformers.modeling_outputs import BaseModelOutputWithPooling
+
+from selva_core.ext.synchformer.hf_src.modeling_ast import ASTForAudioClassification, ASTConfig
+from selva_core.ext.synchformer.motionformer import (AveragePooling, BaseEncoderLayer,
+                                                               TemporalTransformerEncoderLayer)
+from selva_core.ext.synchformer.utils import check_if_file_exists_else_download
+
+
+class AST(torch.nn.Module):
+    def __init__(self,
+                 extract_features: bool = False,
+                 ckpt_path: str = None,
+                 feat_type: str = None,
+                 max_spec_t: int = None,
+                 factorize_freq_time: bool = None,
+                 agg_freq_module: str = None,
+                 agg_time_module: str = None,
+                 add_global_repr: bool = True,
+                 agg_segments_module: str = None,
+                 max_segments: int = None,
+                 ) -> None:
+        '''
+            extract_features: if True, then the model will return the features instead of head's output
+            ckpt_path: is not a path to a ckpt file, but a name of a model from the HuggingFace model hub.
+            feat_type: if extract_features is True, this parameter specifies the type of features to return
+            max_spec_t: if specified, then the model (pos emb) will be patched to support this length of spec
+            factorize_freq_time: if True, then the model will use a factorized freq/time aggregation
+            agg_freq_module: if specified, then the model will use this module for freq aggregation
+            agg_time_module: if specified, then the model will use this module for time aggregation
+            add_global_repr: if True, adds a global representation to the features (aggregation on segments)
+            agg_segments_module: if specified, then the model will use this module for segments aggregation
+            max_segments: if specified, the initialization of PE in the global agg module will use this value.
+                          This should correspond to the max number of segments per video (if None, 16 is used)
+        '''
+        super().__init__()
+        self.extract_features = extract_features
+        self.ckpt_path = ckpt_path
+        self.max_spec_t = max_spec_t
+        self.max_segments = max_segments
+
+        # depending on whether the feat extractor was pre-trained contrastively or not, we need to
+        # load the state dict differently.
+
+        # if ckpt is specified, then load the model from the HuggingFace model hub, otherwise init a new model
+        if ckpt_path == 'MIT/ast-finetuned-audioset-10-10-0.4593':
+            revision = 'c1c0c66'  # fixing the revision for compatibility (V4.27.4)
+            self.config = ASTConfig.from_pretrained(ckpt_path, revision=revision)
+            full_model = ASTForAudioClassification.from_pretrained(ckpt_path, revision=revision)
+            logging.info(f'Loaded AST from {ckpt_path}')
+        else:
+            self.config = ASTConfig()
+            self.config.num_labels = 527  # 2 by default, audioset has 527 labels
+            full_model = ASTForAudioClassification(self.config)
+            logging.info('Initialized AST from scratch with the AST AudioSet config')
+
+        was_pt_on_avclip = ckpt_path is not None and ckpt_path.endswith('.pt')
+
+        # feature extractor
+        self.ast = full_model.audio_spectrogram_transformer
+
+        if self.extract_features:
+            # assign `feat_type` (use default if not specified)
+            self.feat_type = 'last_hidden_state' if feat_type is None else feat_type
+            # define adapters if needed
+            self.factorize_freq_time = factorize_freq_time
+            # avoiding code duplication (used only if agg_*_module is TransformerEncoderLayer)
+            transf_enc_layer_kwargs = dict(
+                d_model=self.config.hidden_size, nhead=self.config.num_attention_heads,
+                dim_feedforward=self.config.intermediate_size, activation=nn.GELU(), batch_first=True,
+                dropout=self.config.attention_probs_dropout_prob, layer_norm_eps=1e-6, norm_first=True,
+            )
+            if factorize_freq_time:
+                self.feat_type = 'last_hidden_state'  # this feat_type supports factorization
+                # frequency aggreration
+                if agg_freq_module == 'TransformerEncoderLayer':
+                    self.freq_attn_agg = FrequencyTransformerEncoderLayer(**transf_enc_layer_kwargs)
+                elif agg_freq_module == 'AveragePooling':
+                    self.freq_attn_agg = AveragePooling(avg_pattern='BS D f t -> BS D t',
+                                                        then_permute_pattern='BS D t -> BS t D')
+                # time aggreration
+                if agg_time_module == 'TransformerEncoderLayer':
+                    self.temp_attn_agg = TemporalTransformerEncoderLayer(**transf_enc_layer_kwargs)
+                elif agg_time_module == 'AveragePooling':
+                    self.temp_attn_agg = AveragePooling(avg_pattern='BS t D -> BS D')
+                elif 'Identity' in agg_time_module:
+                    self.temp_attn_agg = nn.Identity()
+            # define a global aggregation layer (aggregarate over segments)
+            self.add_global_repr = add_global_repr
+            if add_global_repr:
+                if agg_segments_module == 'TransformerEncoderLayer':
+                    # we can reuse the same layer as for temporal factorization (B, dim_to_agg, D) -> (B, D)
+                    # we need to add pos emb (PE) because previously we added the same PE for each segment
+                    pos_max_len = max_segments if max_segments is not None else 16  # 16 = 10sec//0.64sec + 1
+                    self.global_attn_agg = TemporalTransformerEncoderLayer(
+                        add_pos_emb=True, pos_emb_drop=self.config.hidden_dropout_prob,
+                        pos_max_len=pos_max_len, **transf_enc_layer_kwargs
+                    )
+                elif agg_segments_module == 'AveragePooling':
+                    self.global_attn_agg = AveragePooling(avg_pattern='B S D -> B D')
+        else:
+            self.classifier = full_model.classifier
+
+        # AST.device fails with AttributeError. This is a workaround
+        self.device = full_model.device
+
+        # pre-trained on 12*101+2=1214 tokens, but we have less (e.g. 12*6+2=74)
+        self.patch_position_emb()
+
+        if was_pt_on_avclip:
+            # we need to filter out the state_dict of the AVCLIP model (has both A and V extractors)
+            # and keep only the state_dict of the feat extractor
+            check_if_file_exists_else_download(self.ckpt_path)
+            ckpt = torch.load(ckpt_path, map_location='cpu')
+            ckpt_weights = dict()
+            for k, v in ckpt['state_dict'].items():
+                if k.startswith(('module.a_encoder.', 'a_encoder.')):
+                    k = k.replace('module.', '').replace('a_encoder.', '')
+                    ckpt_weights[k] = v
+            _load_status = self.load_state_dict(ckpt_weights, strict=False)
+            if len(_load_status.missing_keys) > 0 or len(_load_status.unexpected_keys) > 0:
+                logging.warning(f'Loading exact afeat_extractor ckpt from {self.ckpt_path} failed. \n' \
+                                f'Missing keys ({len(_load_status.missing_keys)}): ' \
+                                f'{_load_status.missing_keys}, \n' \
+                                f'Unexpected keys ({len(_load_status.unexpected_keys)}): ' \
+                                f'{_load_status.unexpected_keys} \n' \
+                                f'temp_attn_agg are expected to be missing if ckpt was pt contrastively.')
+            else:
+                logging.info(f'Loading afeat_extractor ckpt from {self.ckpt_path} succeeded.')
+
+        # print the number of parameters
+        logging.info(f'AST: {sum(p.numel() for p in self.parameters() if p.requires_grad):,}')
+
+    def forward(self, x: torch.Tensor, for_loop: bool = False, cont_mask: torch.Tensor = None,
+                **ast_kwargs) -> torch.Tensor:
+        '''
+            x: (B, S, T, F) where S is number of segments, F is number of (mel) frequency bins,
+            ast_kwargs: additional arguments for the AST model
+            cont_mask: (B, S, T, F) where 0s are the values to be masked out
+            if `for_loop=True`, we use a for loop to extract features for each segment separately.
+            if `for_loop=False`, we extract features for all segments at once.
+                Using the for loop is slower but more memory efficient, while using all segments at once
+                is faster but more memory inefficient.
+                Using for loop allows to control the memory footprint by varying the number of videos in a
+                batch (batch size) rather than the number of segments in a video.
+        '''
+        B, S, T, F = x.shape
+
+        if for_loop:
+            assert cont_mask is None, 'cont_mask is not supported with for_loop=True'
+            orig_shape_s = (B, 1, T, F)
+            # NOTE: since x is (B, S, T, F), and forward_segments expects (BS, T, F).
+            # (B, S, T, F)[:, s] is (B, T, F) or (BS, T, F) if S=1.
+            x = torch.cat(
+                [self.forward_segments(x[:, s], orig_shape_s, **ast_kwargs).unsqueeze(1) for s in range(S)],
+                dim=1)
+        else:
+            orig_shape = (B, S, T, F)
+            x = x.view(B * S, T, F)
+            if cont_mask is not None:
+                cont_mask = cont_mask.reshape(B * S, T, F)
+            # AST expects a tensor of shape (B*S, T, F).
+            x = self.forward_segments(x, orig_shape=orig_shape, cont_mask=cont_mask, **ast_kwargs)
+            # unpack the segments (using rest dimensions to support different shapes e.g. (BS, D) or (BS, t, D))
+            x = x.view(B, S, *x.shape[1:])
+        # x now is of shape (B, S, D) or (B, S, t, D) if `self.temp_attn_agg` is `Identity`
+
+        global_x = None
+        if self.extract_features and self.add_global_repr:  # lazy execution, throws AttributeError
+            assert len(x.shape) == 3, f'Local representation should be (B, S, D) {x.shape}'
+            global_x = self.global_attn_agg(x)  # (B, D)
+
+        return x, global_x  # x is (B, S, ...), global_x is (B, D) or None
+
+    def forward_segments(self, x, orig_shape: tuple, cont_mask: torch.Tensor = None, **ast_kwargs):
+        '''x is (BS, T, F), where S is the number of segments; cont_mask is (BS, T, F): 0s to be masked out'''
+        # 'pooler_output': (B, D); or 'last_hidden_state: (B, T, D) where T is [CLS, DISTILL, <tokens>]
+        # x_mask is (B, T) where 0s are the values to be masked out
+        x, x_mask = self.ast(x, cont_mask=cont_mask, **ast_kwargs)
+
+        if self.extract_features:
+            x = self.get_features_by_type(x)
+            if self.factorize_freq_time:
+                x = self.restore_freq_temp_dims(x, orig_shape)  # (BS, D, f, t) <- (B*S, T, D)
+                if cont_mask is not None:
+                    # duplicating the mask for the latent dimension (D) to be compatible with the next func
+                    x_mask = x_mask.unsqueeze(-1).expand(-1, -1, self.config.hidden_size)
+                    x_mask = self.restore_freq_temp_dims(x_mask, orig_shape)  # (BS, D, f, t) <- (B*S, T, D)
+                    # again removing the latent
+                    x_mask = x_mask[:, 0, :, :]
+                else:
+                    x_mask = None
+                x = self.freq_attn_agg(x, x_mask)  # (BS, t, D)
+                x = self.temp_attn_agg(x)  # (BS, D) or (BS, t, D) if self.temp_attn_agg is Identity
+        else:
+            x = x['pooler_output']
+            x = self.classifier(x)
+        return x
+
+    def get_features_by_type(self, x: BaseModelOutputWithPooling) -> torch.Tensor:
+        if self.feat_type == 'pooler_output':
+            return x['pooler_output']  # (B, D)
+        elif self.feat_type == 'CLS':
+            return x['last_hidden_state'][:, 0, :]  # (B, D)
+        elif self.feat_type == 'last_hidden_state':
+            return x['last_hidden_state']  # (B, 2+T, D)
+        elif self.feat_type == 'last_hidden_state_no_AUX':
+            return x['last_hidden_state'][:, 2:, :]  # (B, T, D) removing CLS and distill tokens
+        else:
+            raise ValueError(f'Unknown feature type: {self.feat_type}')
+
+    def restore_freq_temp_dims(self, feats, orig_shape: tuple):
+        '''
+            feats are of shape (B*S, T, D)
+                where T = 2 + f * t (if feat_type == 'last_hidden_state')
+                where T =     f * t (if feat_type == 'last_hidden_state_no_AUX')
+            Our goal is to make them of shape (B*S, f, t, D) where f and t are dimensions after patching.
+            From `self.ast.embeddings.patch_embeddings`, it follows that we could reshape feats:
+                `feats.transpose(1, 2).view(B*S, D, f, t)`
+
+            (Similar function is defined in for RGB features in `motionformer.py`)
+        '''
+        B, S, T, F = orig_shape
+        D = self.config.hidden_size
+
+        # num patches in each dimension
+        f, t = self.ast.embeddings.get_shape(self.config)
+
+        if self.feat_type == 'last_hidden_state':
+            feats = feats[:, 2:, :]  # removing CLS and distill tokens
+
+        feats = feats.permute(0, 2, 1)  # (B*S, D, T)
+        feats = feats.view(B * S, D, f, t)  # (B*S, D, f, t)
+
+        return feats
+
+    def patch_position_emb(self):
+        if self.max_spec_t is not None:
+            self.config.max_length = self.max_spec_t
+        f, t = self.ast.embeddings.get_shape(self.config)
+        shortened = self.ast.embeddings.position_embeddings[:, :f*t+2].clone()  # +2 for CLS and distill tokens
+        self.ast.embeddings.position_embeddings = torch.nn.Parameter(shortened).to(self.device)
+
+    def to(self, device):
+        '''AST.device fails with AttributeError. This is a workaround. '''
+        self.device = torch.device(device)
+        return super().to(device)
+
+
+class FrequencyTransformerEncoderLayer(BaseEncoderLayer):
+    ''' This layer is used to aggregate the features along the frequency axis.
+    It follows the same logic as spatio-temporal aggregation in visual feature extractor.
+    Thus, it is recommended to check the definition of `BaseEncoderLayer` in `motionformer.py` '''
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor = None) -> torch.Tensor:
+        ''' x: (B*S, D, f, t); if specified x_mask (B*S, f, t), 0s are the values to be masked out '''
+        BS, D, f, t = x.shape
+
+        # time as a batch dimension
+        x = x.permute(0, 3, 2, 1)  # (B*S, t, f, D)
+        x = x.reshape(BS * t, f, D)  # .view() fails with non-contiguous memory
+        # similar to mask
+        if x_mask is not None:
+            x_mask = x_mask.permute(0, 2, 1)  # (B*S, t, f)
+            x_mask = x_mask.reshape(BS * t, f)
+
+        # apply encoder layer (BaseEncoderLayer.forward) - it will add CLS token and output its representation
+        x = super().forward(x=x, x_mask=x_mask)  # (B*S*t, D)
+
+        # reshape back to (B*S, t, D)
+        x = x.view(BS, t, D)
+
+        return x  # (B*S, t, D)

selva_core/ext/synchformer/divided_224_16x4.yaml

@@ -0,0 +1,84 @@
+TRAIN:
+  ENABLE: True
+  DATASET: Ssv2
+  BATCH_SIZE: 32
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+  CHECKPOINT_EPOCH_RESET: True
+  CHECKPOINT_FILE_PATH: /checkpoint/fmetze/neurips_sota/40944587/checkpoints/checkpoint_epoch_00035.pyth
+DATA:
+  NUM_FRAMES: 16
+  SAMPLING_RATE: 4
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+  MEAN: [0.5, 0.5, 0.5]
+  STD: [0.5, 0.5, 0.5]
+  PATH_TO_DATA_DIR: /private/home/mandelapatrick/slowfast/data/ssv2
+  PATH_PREFIX: /datasets01/SomethingV2/092720/20bn-something-something-v2-frames
+  INV_UNIFORM_SAMPLE: True
+  RANDOM_FLIP: False
+  REVERSE_INPUT_CHANNEL: True
+  USE_RAND_AUGMENT: True
+  RE_PROB: 0.0
+  USE_REPEATED_AUG: False
+  USE_RANDOM_RESIZE_CROPS: False
+  COLORJITTER: False
+  GRAYSCALE: False
+  GAUSSIAN: False
+SOLVER:
+  BASE_LR: 1e-4
+  LR_POLICY: steps_with_relative_lrs
+  LRS: [1, 0.1, 0.01]
+  STEPS: [0, 20, 30]
+  MAX_EPOCH: 35
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 5e-2
+  WARMUP_EPOCHS: 0.0
+  OPTIMIZING_METHOD: adamw
+  USE_MIXED_PRECISION: True
+  SMOOTHING: 0.2
+SLOWFAST:
+  ALPHA: 8
+VIT:
+  PATCH_SIZE: 16
+  PATCH_SIZE_TEMP: 2
+  CHANNELS: 3
+  EMBED_DIM: 768
+  DEPTH: 12
+  NUM_HEADS: 12
+  MLP_RATIO: 4
+  QKV_BIAS: True
+  VIDEO_INPUT: True
+  TEMPORAL_RESOLUTION: 8
+  USE_MLP: True
+  DROP: 0.0
+  POS_DROPOUT: 0.0
+  DROP_PATH: 0.2
+  IM_PRETRAINED: True
+  HEAD_DROPOUT: 0.0
+  HEAD_ACT: tanh
+  PRETRAINED_WEIGHTS: vit_1k
+  ATTN_LAYER: divided
+MODEL:
+  NUM_CLASSES: 174
+  ARCH: slow
+  MODEL_NAME: VisionTransformer
+  LOSS_FUNC: cross_entropy
+TEST:
+  ENABLE: True
+  DATASET: Ssv2
+  BATCH_SIZE: 64
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 4
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 4
+RNG_SEED: 0
+OUTPUT_DIR: .
+TENSORBOARD:
+  ENABLE: True

selva_core/ext/synchformer/hf_src/modeling_ast.py

@@ -0,0 +1,662 @@
+# coding=utf-8
+# Copyright 2022 MIT and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified by v-iashin to support token masking
+
+""" PyTorch Audio Spectrogram Transformer (AST) model."""
+
+import math
+from typing import Dict, List, Optional, Set, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, SequenceClassifierOutput
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
+from transformers.models.audio_spectrogram_transformer.modeling_audio_spectrogram_transformer import ASTConfig
+from transformers.utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "ASTConfig"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "MIT/ast-finetuned-audioset-10-10-0.4593"
+_EXPECTED_OUTPUT_SHAPE = [1, 1214, 768]
+
+# Audio classification docstring
+_SEQ_CLASS_CHECKPOINT = "MIT/ast-finetuned-audioset-10-10-0.4593"
+_SEQ_CLASS_EXPECTED_OUTPUT = "'Speech'"
+_SEQ_CLASS_EXPECTED_LOSS = 0.17
+
+
+AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "MIT/ast-finetuned-audioset-10-10-0.4593",
+    # See all Audio Spectrogram Transformer models at https://huggingface.co/models?filter=ast
+]
+
+
+class ASTEmbeddings(nn.Module):
+    """
+    Construct the CLS token, position and patch embeddings.
+    """
+
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.distillation_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.patch_embeddings = ASTPatchEmbeddings(config)
+
+        frequency_out_dimension, time_out_dimension = self.get_shape(config)
+        num_patches = frequency_out_dimension * time_out_dimension
+        self.position_embeddings = nn.Parameter(torch.zeros(1, num_patches + 2, config.hidden_size))
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def get_shape(self, config):
+        # see Karpathy's cs231n blog on how to calculate the output dimensions
+        # https://cs231n.github.io/convolutional-networks/#conv
+        frequency_out_dimension = (config.num_mel_bins - config.patch_size) // config.frequency_stride + 1
+        time_out_dimension = (config.max_length - config.patch_size) // config.time_stride + 1
+
+        return frequency_out_dimension, time_out_dimension
+
+    def forward(self, input_values: torch.Tensor) -> torch.Tensor:
+        batch_size = input_values.shape[0]
+        embeddings = self.patch_embeddings(input_values)
+
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
+        distillation_tokens = self.distillation_token.expand(batch_size, -1, -1)
+        embeddings = torch.cat((cls_tokens, distillation_tokens, embeddings), dim=1)
+        embeddings = embeddings + self.position_embeddings
+        embeddings = self.dropout(embeddings)
+
+        return embeddings
+
+
+class ASTPatchEmbeddings(nn.Module):
+    """
+    This class turns `input_values` into the initial `hidden_states` (patch embeddings) of shape `(batch_size,
+    seq_length, hidden_size)` to be consumed by a Transformer.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+
+        patch_size = config.patch_size
+        frequency_stride = config.frequency_stride
+        time_stride = config.time_stride
+
+        self.projection = nn.Conv2d(
+            1, config.hidden_size, kernel_size=(patch_size, patch_size), stride=(frequency_stride, time_stride)
+        )
+
+    def forward(self, input_values: torch.Tensor) -> torch.Tensor:
+        input_values = input_values.unsqueeze(1)
+        input_values = input_values.transpose(2, 3)
+        embeddings = self.projection(input_values).flatten(2).transpose(1, 2)
+        return embeddings
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfAttention with ViT->AST
+class ASTSelfAttention(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size {config.hidden_size,} is not a multiple of the number of attention "
+                f"heads {config.num_attention_heads}."
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self, hidden_states, tok_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        # apply masking if provided, tok_mask is (BS, N): 1s - keep; attention_scores is (BS, H, N, N)
+        if tok_mask is not None:
+            BS, N = tok_mask.shape
+            attention_scores = attention_scores.masked_fill(tok_mask.view(BS, 1, 1, N) == 0, float('-inf'))
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfOutput with ViT->AST
+class ASTSelfOutput(nn.Module):
+    """
+    The residual connection is defined in ASTLayer instead of here (as is the case with other models), due to the
+    layernorm applied before each block.
+    """
+
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTAttention with ViT->AST
+class ASTAttention(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.attention = ASTSelfAttention(config)
+        self.output = ASTSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads: Set[int]) -> None:
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.attention.query = prune_linear_layer(self.attention.query, index)
+        self.attention.key = prune_linear_layer(self.attention.key, index)
+        self.attention.value = prune_linear_layer(self.attention.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
+        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        tok_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_outputs = self.attention(hidden_states, tok_mask, head_mask, output_attentions)
+
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTIntermediate with ViT->AST
+class ASTIntermediate(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTOutput with ViT->AST
+class ASTOutput(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        hidden_states = hidden_states + input_tensor
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTLayer with ViT->AST
+class ASTLayer(nn.Module):
+    """This corresponds to the Block class in the timm implementation."""
+
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = ASTAttention(config)
+        self.intermediate = ASTIntermediate(config)
+        self.output = ASTOutput(config)
+        self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        tok_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        self_attention_outputs = self.attention(
+            self.layernorm_before(hidden_states),  # in AST, layernorm is applied before self-attention
+            tok_mask,
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        # first residual connection
+        hidden_states = attention_output + hidden_states
+
+        # in AST, layernorm is also applied after self-attention
+        layer_output = self.layernorm_after(hidden_states)
+        layer_output = self.intermediate(layer_output)
+
+        # second residual connection is done here
+        layer_output = self.output(layer_output, hidden_states)
+
+        outputs = (layer_output,) + outputs
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTEncoder with ViT->AST
+class ASTEncoder(nn.Module):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([ASTLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        tok_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ) -> Union[tuple, BaseModelOutput]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    tok_mask,
+                    layer_head_mask,
+                )
+            else:
+                layer_outputs = layer_module(hidden_states, tok_mask, layer_head_mask, output_attentions)
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ASTPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ASTConfig
+    base_model_prefix = "audio_spectrogram_transformer"
+    main_input_name = "input_values"
+    supports_gradient_checkpointing = True
+
+    # Copied from transformers.models.deit.modeling_deit.DeiTPreTrainedModel._init_weights
+    def _init_weights(self, module: Union[nn.Linear, nn.Conv2d, nn.LayerNorm]) -> None:
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            # Upcast the input in `fp32` and cast it back to desired `dtype` to avoid
+            # `trunc_normal_cpu` not implemented in `half` issues
+            module.weight.data = nn.init.trunc_normal_(
+                module.weight.data.to(torch.float32), mean=0.0, std=self.config.initializer_range
+            ).to(module.weight.dtype)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    # Copied from transformers.models.vit.modeling_vit.ViTPreTrainedModel._set_gradient_checkpointing with ViT->AST
+    def _set_gradient_checkpointing(self, module: ASTEncoder, value: bool = False) -> None:
+        if isinstance(module, ASTEncoder):
+            module.gradient_checkpointing = value
+
+
+AUDIO_SPECTROGRAM_TRANSFORMER_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it
+    as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`ASTConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+AUDIO_SPECTROGRAM_TRANSFORMER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoFeatureExtractor`]. See
+            [`ASTFeatureExtractor.__call__`] for details.
+
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare AST Model transformer outputting raw hidden-states without any specific head on top.",
+    AUDIO_SPECTROGRAM_TRANSFORMER_START_DOCSTRING,
+)
+class ASTModel(ASTPreTrainedModel):
+    def __init__(self, config: ASTConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = ASTEmbeddings(config)
+        self.encoder = ASTEncoder(config)
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> ASTPatchEmbeddings:
+        return self.embeddings.patch_embeddings
+
+    def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(AUDIO_SPECTROGRAM_TRANSFORMER_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+        modality="audio",
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self,
+        input_values: Optional[torch.Tensor] = None,
+        cont_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_values is None:
+            raise ValueError("You have to specify input_values")
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(input_values)
+
+        # transforms the mask that has spectrogram dims to the token masking which is obtained after patching.
+        # Due to the ovelap in patching, getting the token mask from spectrogram mask is not straightforward,
+        # because one 16x16 content patch is encoded in two tokens if stride is <16. So, to get the mask for
+        # tokens I will apply the patching func (self.embeddings) to the tensor with infinities at the masked
+        # content position. For infs, the patching fn will return nans, which I'll use to get the token mask.
+        if cont_mask is not None:
+            indicator = torch.ones_like(input_values).to(input_values.dtype)
+            # replace content mask (0s) with infs
+            indicator[~cont_mask] = torch.inf
+            # apply patching; now nans are where the content mask was
+            with torch.no_grad():
+                indicator = self.embeddings(indicator)  # BS, N, D
+            # replace nans with 0s; these are the tokens that correspond to the masked content
+            tok_mask = ~torch.isnan(indicator)
+            # since all values in the D-dimension (latent) will also be nans, we can just use the first el
+            tok_mask = tok_mask[:, :, 0]   # (BS, 2+num_patches) -- 2 is from CLS and DISTIL tokens
+        else:
+            tok_mask = None
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            tok_mask=tok_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        sequence_output = self.layernorm(sequence_output)
+
+        pooled_output = (sequence_output[:, 0] + sequence_output[:, 1]) / 2
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        ), tok_mask
+
+
+class ASTMLPHead(nn.Module):
+    def __init__(self, config: ASTConfig):
+        super().__init__()
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dense = nn.Linear(
+            config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
+
+    def forward(self, hidden_state):
+        hidden_state = self.layernorm(hidden_state)
+        hidden_state = self.dense(hidden_state)
+        return hidden_state
+
+
+@add_start_docstrings(
+    """
+    Audio Spectrogram Transformer model with an audio classification head on top (a linear layer on top of the pooled
+    output) e.g. for datasets like AudioSet, Speech Commands v2.
+    """,
+    AUDIO_SPECTROGRAM_TRANSFORMER_START_DOCSTRING,
+)
+class ASTForAudioClassification(ASTPreTrainedModel):
+    def __init__(self, config: ASTConfig) -> None:
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        self.audio_spectrogram_transformer = ASTModel(config)
+
+        # Classifier head
+        self.classifier = ASTMLPHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(AUDIO_SPECTROGRAM_TRANSFORMER_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_SEQ_CLASS_CHECKPOINT,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        modality="audio",
+        expected_output=_SEQ_CLASS_EXPECTED_OUTPUT,
+        expected_loss=_SEQ_CLASS_EXPECTED_LOSS,
+    )
+    def forward(
+        self,
+        input_values: Optional[torch.Tensor] = None,
+        cont_mask: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[tuple, SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the audio classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.audio_spectrogram_transformer(
+            input_values,
+            cont_mask=cont_mask,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

selva_core/ext/synchformer/motionformer.py

@@ -0,0 +1,400 @@
+import logging
+from pathlib import Path
+
+import einops
+import torch
+from omegaconf import OmegaConf
+from timm.layers import trunc_normal_
+from torch import nn
+
+from selva_core.ext.synchformer.utils import check_if_file_exists_else_download
+from selva_core.ext.synchformer.video_model_builder import VisionTransformer
+
+FILE2URL = {
+    # cfg
+    'motionformer_224_16x4.yaml':
+    'https://raw.githubusercontent.com/facebookresearch/Motionformer/bf43d50/configs/SSV2/motionformer_224_16x4.yaml',
+    'joint_224_16x4.yaml':
+    'https://raw.githubusercontent.com/facebookresearch/Motionformer/bf43d50/configs/SSV2/joint_224_16x4.yaml',
+    'divided_224_16x4.yaml':
+    'https://raw.githubusercontent.com/facebookresearch/Motionformer/bf43d50/configs/SSV2/divided_224_16x4.yaml',
+    # ckpt
+    'ssv2_motionformer_224_16x4.pyth':
+    'https://dl.fbaipublicfiles.com/motionformer/ssv2_motionformer_224_16x4.pyth',
+    'ssv2_joint_224_16x4.pyth':
+    'https://dl.fbaipublicfiles.com/motionformer/ssv2_joint_224_16x4.pyth',
+    'ssv2_divided_224_16x4.pyth':
+    'https://dl.fbaipublicfiles.com/motionformer/ssv2_divided_224_16x4.pyth',
+}
+
+
+class MotionFormer(VisionTransformer):
+    ''' This class serves three puposes:
+            1. Renames the class to MotionFormer.
+            2. Downloads the cfg from the original repo and patches it if needed.
+            3. Takes care of feature extraction by redefining .forward()
+                - if `extract_features=True` and `factorize_space_time=False`,
+                    the output is of shape (B, T, D) where T = 1 + (224 // 16) * (224 // 16) * 8
+                - if `extract_features=True` and `factorize_space_time=True`, the output is of shape (B*S, D)
+                    and spatial and temporal transformer encoder layers are used.
+                - if `extract_features=True` and `factorize_space_time=True` as well as `add_global_repr=True`
+                    the output is of shape (B, D) and spatial and temporal transformer encoder layers
+                    are used as well as the global representation is extracted from segments (extra pos emb
+                    is added).
+    '''
+
+    def __init__(
+        self,
+        extract_features: bool = False,
+        ckpt_path: str = None,
+        factorize_space_time: bool = None,
+        agg_space_module: str = None,
+        agg_time_module: str = None,
+        add_global_repr: bool = True,
+        agg_segments_module: str = None,
+        max_segments: int = None,
+    ):
+        self.extract_features = extract_features
+        self.ckpt_path = ckpt_path
+        self.factorize_space_time = factorize_space_time
+
+        if self.ckpt_path is not None:
+            check_if_file_exists_else_download(self.ckpt_path, FILE2URL)
+            ckpt = torch.load(self.ckpt_path, map_location='cpu')
+            mformer_ckpt2cfg = {
+                'ssv2_motionformer_224_16x4.pyth': 'motionformer_224_16x4.yaml',
+                'ssv2_joint_224_16x4.pyth': 'joint_224_16x4.yaml',
+                'ssv2_divided_224_16x4.pyth': 'divided_224_16x4.yaml',
+            }
+            # init from motionformer ckpt or from our Stage I ckpt
+            # depending on whether the feat extractor was pre-trained on AVCLIPMoCo or not, we need to
+            # load the state dict differently
+            was_pt_on_avclip = self.ckpt_path.endswith(
+                '.pt')  # checks if it is a stage I ckpt (FIXME: a bit generic)
+            if self.ckpt_path.endswith(tuple(mformer_ckpt2cfg.keys())):
+                cfg_fname = mformer_ckpt2cfg[Path(self.ckpt_path).name]
+            elif was_pt_on_avclip:
+                # TODO: this is a hack, we should be able to get the cfg from the ckpt (earlier ckpt didn't have it)
+                s1_cfg = ckpt.get('args', None)  # Stage I cfg
+                if s1_cfg is not None:
+                    s1_vfeat_extractor_ckpt_path = s1_cfg.model.params.vfeat_extractor.params.ckpt_path
+                    # if the stage I ckpt was initialized from a motionformer ckpt or train from scratch
+                    if s1_vfeat_extractor_ckpt_path is not None:
+                        cfg_fname = mformer_ckpt2cfg[Path(s1_vfeat_extractor_ckpt_path).name]
+                    else:
+                        cfg_fname = 'divided_224_16x4.yaml'
+                else:
+                    cfg_fname = 'divided_224_16x4.yaml'
+            else:
+                raise ValueError(f'ckpt_path {self.ckpt_path} is not supported.')
+        else:
+            was_pt_on_avclip = False
+            cfg_fname = 'divided_224_16x4.yaml'
+            # logging.info(f'No ckpt_path provided, using {cfg_fname} config.')
+
+        if cfg_fname in ['motionformer_224_16x4.yaml', 'divided_224_16x4.yaml']:
+            pos_emb_type = 'separate'
+        elif cfg_fname == 'joint_224_16x4.yaml':
+            pos_emb_type = 'joint'
+
+        self.mformer_cfg_path = Path(__file__).absolute().parent / cfg_fname
+
+        check_if_file_exists_else_download(self.mformer_cfg_path, FILE2URL)
+        mformer_cfg = OmegaConf.load(self.mformer_cfg_path)
+        logging.info(f'Loading MotionFormer config from {self.mformer_cfg_path.absolute()}')
+
+        # patch the cfg (from the default cfg defined in the repo `Motionformer/slowfast/config/defaults.py`)
+        mformer_cfg.VIT.ATTN_DROPOUT = 0.0
+        mformer_cfg.VIT.POS_EMBED = pos_emb_type
+        mformer_cfg.VIT.USE_ORIGINAL_TRAJ_ATTN_CODE = True
+        mformer_cfg.VIT.APPROX_ATTN_TYPE = 'none'  # guessing
+        mformer_cfg.VIT.APPROX_ATTN_DIM = 64  # from ckpt['cfg']
+
+        # finally init VisionTransformer with the cfg
+        super().__init__(mformer_cfg)
+
+        # load the ckpt now if ckpt is provided and not from AVCLIPMoCo-pretrained ckpt
+        if (self.ckpt_path is not None) and (not was_pt_on_avclip):
+            _ckpt_load_status = self.load_state_dict(ckpt['model_state'], strict=False)
+            if len(_ckpt_load_status.missing_keys) > 0 or len(
+                    _ckpt_load_status.unexpected_keys) > 0:
+                logging.warning(f'Loading exact vfeat_extractor ckpt from {self.ckpt_path} failed.' \
+                                f'Missing keys: {_ckpt_load_status.missing_keys}, ' \
+                                f'Unexpected keys: {_ckpt_load_status.unexpected_keys}')
+            else:
+                logging.info(f'Loading vfeat_extractor ckpt from {self.ckpt_path} succeeded.')
+
+        if self.extract_features:
+            assert isinstance(self.norm,
+                              nn.LayerNorm), 'early x[:, 1:, :] may not be safe for per-tr weights'
+            # pre-logits are Sequential(nn.Linear(emb, emd), act) and `act` is tanh but see the logger
+            self.pre_logits = nn.Identity()
+            # we don't need the classification head (saving memory)
+            self.head = nn.Identity()
+            self.head_drop = nn.Identity()
+            # avoiding code duplication (used only if agg_*_module is TransformerEncoderLayer)
+            transf_enc_layer_kwargs = dict(
+                d_model=self.embed_dim,
+                nhead=self.num_heads,
+                activation=nn.GELU(),
+                batch_first=True,
+                dim_feedforward=self.mlp_ratio * self.embed_dim,
+                dropout=self.drop_rate,
+                layer_norm_eps=1e-6,
+                norm_first=True,
+            )
+            # define adapters if needed
+            if self.factorize_space_time:
+                if agg_space_module == 'TransformerEncoderLayer':
+                    self.spatial_attn_agg = SpatialTransformerEncoderLayer(
+                        **transf_enc_layer_kwargs)
+                elif agg_space_module == 'AveragePooling':
+                    self.spatial_attn_agg = AveragePooling(avg_pattern='BS D t h w -> BS D t',
+                                                           then_permute_pattern='BS D t -> BS t D')
+                if agg_time_module == 'TransformerEncoderLayer':
+                    self.temp_attn_agg = TemporalTransformerEncoderLayer(**transf_enc_layer_kwargs)
+                elif agg_time_module == 'AveragePooling':
+                    self.temp_attn_agg = AveragePooling(avg_pattern='BS t D -> BS D')
+                elif 'Identity' in agg_time_module:
+                    self.temp_attn_agg = nn.Identity()
+            # define a global aggregation layer (aggregarate over segments)
+            self.add_global_repr = add_global_repr
+            if add_global_repr:
+                if agg_segments_module == 'TransformerEncoderLayer':
+                    # we can reuse the same layer as for temporal factorization (B, dim_to_agg, D) -> (B, D)
+                    # we need to add pos emb (PE) because previously we added the same PE for each segment
+                    pos_max_len = max_segments if max_segments is not None else 16  # 16 = 10sec//0.64sec + 1
+                    self.global_attn_agg = TemporalTransformerEncoderLayer(
+                        add_pos_emb=True,
+                        pos_emb_drop=mformer_cfg.VIT.POS_DROPOUT,
+                        pos_max_len=pos_max_len,
+                        **transf_enc_layer_kwargs)
+                elif agg_segments_module == 'AveragePooling':
+                    self.global_attn_agg = AveragePooling(avg_pattern='B S D -> B D')
+
+        if was_pt_on_avclip:
+            # we need to filter out the state_dict of the AVCLIP model (has both A and V extractors)
+            # and keep only the state_dict of the feat extractor
+            ckpt_weights = dict()
+            for k, v in ckpt['state_dict'].items():
+                if k.startswith(('module.v_encoder.', 'v_encoder.')):
+                    k = k.replace('module.', '').replace('v_encoder.', '')
+                    ckpt_weights[k] = v
+            _load_status = self.load_state_dict(ckpt_weights, strict=False)
+            if len(_load_status.missing_keys) > 0 or len(_load_status.unexpected_keys) > 0:
+                logging.warning(f'Loading exact vfeat_extractor ckpt from {self.ckpt_path} failed. \n' \
+                                f'Missing keys ({len(_load_status.missing_keys)}): ' \
+                                f'{_load_status.missing_keys}, \n' \
+                                f'Unexpected keys ({len(_load_status.unexpected_keys)}): ' \
+                                f'{_load_status.unexpected_keys} \n' \
+                                f'temp_attn_agg are expected to be missing if ckpt was pt contrastively.')
+            else:
+                logging.info(f'Loading vfeat_extractor ckpt from {self.ckpt_path} succeeded.')
+
+        # patch_embed is not used in MotionFormer, only patch_embed_3d, because cfg.VIT.PATCH_SIZE_TEMP > 1
+        # but it used to calculate the number of patches, so we need to set keep it
+        self.patch_embed.requires_grad_(False)
+
+    def forward(self, x):
+        '''
+        x is of shape (B, S, C, T, H, W) where S is the number of segments.
+        '''
+        # Batch, Segments, Channels, T=frames, Height, Width
+        B, S, C, T, H, W = x.shape
+        # Motionformer expects a tensor of shape (1, B, C, T, H, W).
+        # The first dimension (1) is a dummy dimension to make the input tensor and won't be used:
+        # see `video_model_builder.video_input`.
+        # x = x.unsqueeze(0)  # (1, B, S, C, T, H, W)
+
+        orig_shape = (B, S, C, T, H, W)
+        x = x.view(B * S, C, T, H, W)  # flatten batch and segments
+        x = self.forward_segments(x, orig_shape=orig_shape)
+        # unpack the segments (using rest dimensions to support different shapes e.g. (BS, D) or (BS, t, D))
+        x = x.view(B, S, *x.shape[1:])
+        # x is now of shape (B*S, D) or (B*S, t, D) if `self.temp_attn_agg` is `Identity`
+
+        return x  # x is (B, S, ...)
+
+    def forward_segments(self, x, orig_shape: tuple) -> torch.Tensor:
+        '''x is of shape (1, BS, C, T, H, W) where S is the number of segments.'''
+        x, x_mask = self.forward_features(x)
+
+        assert self.extract_features
+
+        # (BS, T, D) where T = 1 + (224 // 16) * (224 // 16) * 8
+        x = x[:,
+              1:, :]  # without the CLS token for efficiency (should be safe for LayerNorm and FC)
+        x = self.norm(x)
+        x = self.pre_logits(x)
+        if self.factorize_space_time:
+            x = self.restore_spatio_temp_dims(x, orig_shape)  # (B*S, D, t, h, w) <- (B*S, t*h*w, D)
+
+            x = self.spatial_attn_agg(x, x_mask)  # (B*S, t, D)
+            x = self.temp_attn_agg(
+                x)  # (B*S, D) or (BS, t, D) if `self.temp_attn_agg` is `Identity`
+
+        return x
+
+    def restore_spatio_temp_dims(self, feats: torch.Tensor, orig_shape: tuple) -> torch.Tensor:
+        '''
+            feats are of shape (B*S, T, D) where T = 1 + (224 // 16) * (224 // 16) * 8
+            Our goal is to make them of shape (B*S, t, h, w, D) where h, w are the spatial dimensions.
+            From `self.patch_embed_3d`, it follows that we could reshape feats with:
+                `feats.transpose(1, 2).view(B*S, D, t, h, w)`
+        '''
+        B, S, C, T, H, W = orig_shape
+        D = self.embed_dim
+
+        # num patches in each dimension
+        t = T // self.patch_embed_3d.z_block_size
+        h = self.patch_embed_3d.height
+        w = self.patch_embed_3d.width
+
+        feats = feats.permute(0, 2, 1)  # (B*S, D, T)
+        feats = feats.view(B * S, D, t, h, w)  # (B*S, D, t, h, w)
+
+        return feats
+
+
+class BaseEncoderLayer(nn.TransformerEncoderLayer):
+    '''
+        This is a wrapper around nn.TransformerEncoderLayer that adds a CLS token
+        to the sequence and outputs the CLS token's representation.
+        This base class parents both SpatialEncoderLayer and TemporalEncoderLayer for the RGB stream
+        and the FrequencyEncoderLayer and TemporalEncoderLayer for the audio stream stream.
+        We also, optionally, add a positional embedding to the input sequence which
+        allows to reuse it for global aggregation (of segments) for both streams.
+    '''
+
+    def __init__(self,
+                 add_pos_emb: bool = False,
+                 pos_emb_drop: float = None,
+                 pos_max_len: int = None,
+                 *args_transformer_enc,
+                 **kwargs_transformer_enc):
+        super().__init__(*args_transformer_enc, **kwargs_transformer_enc)
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, self.self_attn.embed_dim))
+        trunc_normal_(self.cls_token, std=.02)
+
+        # add positional embedding
+        self.add_pos_emb = add_pos_emb
+        if add_pos_emb:
+            self.pos_max_len = 1 + pos_max_len  # +1 (for CLS)
+            self.pos_emb = nn.Parameter(torch.zeros(1, self.pos_max_len, self.self_attn.embed_dim))
+            self.pos_drop = nn.Dropout(pos_emb_drop)
+            trunc_normal_(self.pos_emb, std=.02)
+
+        self.apply(self._init_weights)
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor = None):
+        ''' x is of shape (B, N, D); if provided x_mask is of shape (B, N)'''
+        batch_dim = x.shape[0]
+
+        # add CLS token
+        cls_tokens = self.cls_token.expand(batch_dim, -1, -1)  # expanding to match batch dimension
+        x = torch.cat((cls_tokens, x), dim=-2)  # (batch_dim, 1+seq_len, D)
+        if x_mask is not None:
+            cls_mask = torch.ones((batch_dim, 1), dtype=torch.bool,
+                                  device=x_mask.device)  # 1=keep; 0=mask
+            x_mask_w_cls = torch.cat((cls_mask, x_mask), dim=-1)  # (batch_dim, 1+seq_len)
+            B, N = x_mask_w_cls.shape
+            # torch expects (N, N) or (B*num_heads, N, N) mask (sadness ahead); torch masks
+            x_mask_w_cls = x_mask_w_cls.reshape(B, 1, 1, N)\
+                                       .expand(-1, self.self_attn.num_heads, N, -1)\
+                                       .reshape(B * self.self_attn.num_heads, N, N)
+            assert x_mask_w_cls.dtype == x_mask_w_cls.bool().dtype, 'x_mask_w_cls.dtype != bool'
+            x_mask_w_cls = ~x_mask_w_cls  # invert mask (1=mask)
+        else:
+            x_mask_w_cls = None
+
+        # add positional embedding
+        if self.add_pos_emb:
+            seq_len = x.shape[
+                1]  # (don't even think about moving it before the CLS token concatenation)
+            assert seq_len <= self.pos_max_len, f'Seq len ({seq_len}) > pos_max_len ({self.pos_max_len})'
+            x = x + self.pos_emb[:, :seq_len, :]
+            x = self.pos_drop(x)
+
+        # apply encoder layer (calls nn.TransformerEncoderLayer.forward);
+        x = super().forward(src=x, src_mask=x_mask_w_cls)  # (batch_dim, 1+seq_len, D)
+
+        # CLS token is expected to hold spatial information for each frame
+        x = x[:, 0, :]  # (batch_dim, D)
+
+        return x
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'cls_token', 'pos_emb'}
+
+
+class SpatialTransformerEncoderLayer(BaseEncoderLayer):
+    ''' Aggregates spatial dimensions by applying attention individually to each frame. '''
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor = None) -> torch.Tensor:
+        ''' x is of shape (B*S, D, t, h, w) where S is the number of segments.
+            if specified x_mask (B*S, t, h, w), 0=masked, 1=kept
+            Returns a tensor of shape (B*S, t, D) pooling spatial information for each frame. '''
+        BS, D, t, h, w = x.shape
+
+        # time as a batch dimension and flatten spatial dimensions as sequence
+        x = einops.rearrange(x, 'BS D t h w -> (BS t) (h w) D')
+        # similar to mask
+        if x_mask is not None:
+            x_mask = einops.rearrange(x_mask, 'BS t h w -> (BS t) (h w)')
+
+        # apply encoder layer (BaseEncoderLayer.forward) - it will add CLS token and output its representation
+        x = super().forward(x=x, x_mask=x_mask)  # (B*S*t, D)
+
+        # reshape back to (B*S, t, D)
+        x = einops.rearrange(x, '(BS t) D -> BS t D', BS=BS, t=t)
+
+        # (B*S, t, D)
+        return x
+
+
+class TemporalTransformerEncoderLayer(BaseEncoderLayer):
+    ''' Aggregates temporal dimension with attention. Also used with pos emb as global aggregation
+    in both streams. '''
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x):
+        ''' x is of shape (B*S, t, D) where S is the number of segments.
+            Returns a tensor of shape (B*S, D) pooling temporal information. '''
+        BS, t, D = x.shape
+
+        # apply encoder layer (BaseEncoderLayer.forward) - it will add CLS token and output its representation
+        x = super().forward(x)  # (B*S, D)
+
+        return x  # (B*S, D)
+
+
+class AveragePooling(nn.Module):
+
+    def __init__(self, avg_pattern: str, then_permute_pattern: str = None) -> None:
+        ''' patterns are e.g. "bs t d -> bs d" '''
+        super().__init__()
+        # TODO: need to register them as buffers (but fails because these are strings)
+        self.reduce_fn = 'mean'
+        self.avg_pattern = avg_pattern
+        self.then_permute_pattern = then_permute_pattern
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor = None) -> torch.Tensor:
+        x = einops.reduce(x, self.avg_pattern, self.reduce_fn)
+        if self.then_permute_pattern is not None:
+            x = einops.rearrange(x, self.then_permute_pattern)
+        return x

selva_core/ext/synchformer/synchformer.py

@@ -0,0 +1,144 @@
+import logging
+from typing import Any, Mapping
+
+import torch
+from torch import nn
+
+from selva_core.ext.synchformer.motionformer import MotionFormer
+from selva_core.ext.synchformer.astransformer import AST
+
+
+class Synchformer(nn.Module):
+
+    def __init__(self, video: bool = True, audio: bool = False):
+        super().__init__()
+        
+        self.video = video
+        self.audio = audio
+        
+        if not video and not audio:
+            raise ValueError('At least one of vis or audio should be True.')
+
+        if self.video:
+            self.vfeat_extractor = MotionFormer(extract_features=True,
+                                                factorize_space_time=True,
+                                                agg_space_module='TransformerEncoderLayer',
+                                                agg_time_module='torch.nn.Identity',
+                                                add_global_repr=False)
+        if self.audio:
+            self.afeat_extractor = AST(extract_features=True,
+                                    max_spec_t=66,
+                                    factorize_freq_time=True,
+                                    agg_freq_module='TransformerEncoderLayer',
+                                    agg_time_module='torch.nn.Identity',
+                                    add_global_repr=False)
+
+        # self.vfeat_extractor = instantiate_from_config(vfeat_extractor)
+        # self.afeat_extractor = instantiate_from_config(afeat_extractor)
+        # # bridging the s3d latent dim (1024) into what is specified in the config
+        # # to match e.g. the transformer dim
+        # self.vproj = instantiate_from_config(vproj)
+        # self.aproj = instantiate_from_config(aproj)
+        # self.transformer = instantiate_from_config(transformer)
+
+    def forward(self, data):
+        video, audio = None, None
+        
+        if self.video and self.audio:
+            video, audio = data
+        elif self.video:
+            video = data
+        elif self.audio:
+            audio = data
+            
+        if self.video and video is not None:
+            video = self.forward_vfeat(video)
+        if self.audio and audio is not None:
+            audio = self.forward_afeat(audio)
+        
+        if self.video and self.audio:
+            return video, audio
+        elif self.video:
+            return video
+        else:
+            return audio
+    
+    def forward_vfeat(self, vis):
+        B, S, Tv, C, H, W = vis.shape
+        vis = vis.permute(0, 1, 3, 2, 4, 5)  # (B, S, C, Tv, H, W)
+        # feat extractors return a tuple of segment-level and global features (ignored for sync)
+        # (B, S, tv, D), e.g. (B, 7, 8, 768)
+        vis = self.vfeat_extractor(vis)
+        return vis
+    
+    def forward_afeat(self, aud):
+        B, S, F, Ta = aud.shape
+        aud = aud.permute(0, 1, 3, 2) # (B, S, Ta, F)
+        aud, _ = self.afeat_extractor(aud)
+        return aud
+        
+
+    def load_state_dict(self, sd: Mapping[str, Any], strict: bool = True):
+        target_keys = (['vfeat_extractor'] if self.video else []) \
+                    + (['afeat_extractor'] if self.audio else [])
+        # discard all entries except vfeat_extractor / afeat_extractor
+        sd = {k: v for k, v in sd.items() if any(k.startswith(tk) 
+                                                 for tk in target_keys)}
+        
+
+        return super().load_state_dict(sd, strict)
+
+
+if __name__ == "__main__":
+    model = Synchformer(video=True, audio=True).cuda().eval()
+    sd = torch.load('/mnt/hdd3/junwon/mmaudio/ext_weights/synchformer_state_dict.pth', weights_only=True)
+    model.load_state_dict(sd)
+
+    vid = torch.randn(2, 7, 16, 3, 224, 224).cuda()
+    features = model.forward_vfeat(vid).detach().cpu()
+    print(features.shape)
+    
+    aud = torch.randn(2, 16000*8).cuda()
+    segment_size = 10_240 # 16000 * (16/25) = 16000 * 0.64
+    step_size = 5_120 # segment_size // 2
+    num_segments = (128000 - segment_size) // step_size + 1
+    segments = []
+    for i in range(num_segments):
+        segments.append(aud[:, i * step_size:i * step_size + segment_size])
+    aud = torch.stack(segments, dim=1) # (B, S, T)
+    print(aud.shape)
+    import torchaudio
+    spec = torchaudio.transforms.MelSpectrogram(
+        sample_rate=16000,
+        win_length=400,
+        hop_length=160,
+        n_fft=1024,
+        n_mels=128,
+    )
+    spec = spec.cuda()
+    aud = spec(aud) # (B, S, F, T)
+    aud = torch.log(aud + 1e-6)
+    max_spec_t = 66
+    if max_spec_t - aud.shape[-1] > 0:
+        # pad the last dim (time) -> (..., n_mels, 0+time+difference)  # safe for batched input
+        pad_dims = (0, max_spec_t - aud.shape[-1])
+        aud = torch.nn.functional.pad(aud, pad_dims, 
+                                        'constant', 0.0)
+    aud = aud[..., :max_spec_t]  # (B, S, F, T=66)
+    MEAN = -4.2677393
+    STD = 4.5689974
+    aud = (aud - MEAN) / (2 * STD)
+    print(aud.shape)
+    
+    from einops import rearrange
+    aud = rearrange(aud, 'b s f t -> (b s) 1 f t')
+    print(aud.shape)
+    aud = model.forward_afeat(aud).detach().cpu()
+    print(aud.shape)
+    aud = rearrange(aud, '(b s) 1 t d -> b (s t) d', b=2)
+    print(aud.shape)
+    
+
+    # extract and save the state dict only
+    # sd = torch.load('./ext_weights/sync_model_audioset.pt')['model']
+    # torch.save(sd, './ext_weights/synchformer_state_dict.pth')

selva_core/ext/synchformer/utils.py

@@ -0,0 +1,92 @@
+from hashlib import md5
+from pathlib import Path
+
+import requests
+from tqdm import tqdm
+
+PARENT_LINK = 'https://a3s.fi/swift/v1/AUTH_a235c0f452d648828f745589cde1219a'
+FNAME2LINK = {
+    # S3: Synchability: AudioSet (run 2)
+    '24-01-22T20-34-52.pt':
+    f'{PARENT_LINK}/sync/sync_models/24-01-22T20-34-52/24-01-22T20-34-52.pt',
+    'cfg-24-01-22T20-34-52.yaml':
+    f'{PARENT_LINK}/sync/sync_models/24-01-22T20-34-52/cfg-24-01-22T20-34-52.yaml',
+    # S2: Synchformer: AudioSet (run 2)
+    '24-01-04T16-39-21.pt':
+    f'{PARENT_LINK}/sync/sync_models/24-01-04T16-39-21/24-01-04T16-39-21.pt',
+    'cfg-24-01-04T16-39-21.yaml':
+    f'{PARENT_LINK}/sync/sync_models/24-01-04T16-39-21/cfg-24-01-04T16-39-21.yaml',
+    # S2: Synchformer: AudioSet (run 1)
+    '23-08-28T11-23-23.pt':
+    f'{PARENT_LINK}/sync/sync_models/23-08-28T11-23-23/23-08-28T11-23-23.pt',
+    'cfg-23-08-28T11-23-23.yaml':
+    f'{PARENT_LINK}/sync/sync_models/23-08-28T11-23-23/cfg-23-08-28T11-23-23.yaml',
+    # S2: Synchformer: LRS3 (run 2)
+    '23-12-23T18-33-57.pt':
+    f'{PARENT_LINK}/sync/sync_models/23-12-23T18-33-57/23-12-23T18-33-57.pt',
+    'cfg-23-12-23T18-33-57.yaml':
+    f'{PARENT_LINK}/sync/sync_models/23-12-23T18-33-57/cfg-23-12-23T18-33-57.yaml',
+    # S2: Synchformer: VGS (run 2)
+    '24-01-02T10-00-53.pt':
+    f'{PARENT_LINK}/sync/sync_models/24-01-02T10-00-53/24-01-02T10-00-53.pt',
+    'cfg-24-01-02T10-00-53.yaml':
+    f'{PARENT_LINK}/sync/sync_models/24-01-02T10-00-53/cfg-24-01-02T10-00-53.yaml',
+    # SparseSync: ft VGGSound-Full
+    '22-09-21T21-00-52.pt':
+    f'{PARENT_LINK}/sync/sync_models/22-09-21T21-00-52/22-09-21T21-00-52.pt',
+    'cfg-22-09-21T21-00-52.yaml':
+    f'{PARENT_LINK}/sync/sync_models/22-09-21T21-00-52/cfg-22-09-21T21-00-52.yaml',
+    # SparseSync: ft VGGSound-Sparse
+    '22-07-28T15-49-45.pt':
+    f'{PARENT_LINK}/sync/sync_models/22-07-28T15-49-45/22-07-28T15-49-45.pt',
+    'cfg-22-07-28T15-49-45.yaml':
+    f'{PARENT_LINK}/sync/sync_models/22-07-28T15-49-45/cfg-22-07-28T15-49-45.yaml',
+    # SparseSync: only pt on LRS3
+    '22-07-13T22-25-49.pt':
+    f'{PARENT_LINK}/sync/sync_models/22-07-13T22-25-49/22-07-13T22-25-49.pt',
+    'cfg-22-07-13T22-25-49.yaml':
+    f'{PARENT_LINK}/sync/sync_models/22-07-13T22-25-49/cfg-22-07-13T22-25-49.yaml',
+    # SparseSync: feature extractors
+    'ResNetAudio-22-08-04T09-51-04.pt':
+    f'{PARENT_LINK}/sync/ResNetAudio-22-08-04T09-51-04.pt',  # 2s
+    'ResNetAudio-22-08-03T23-14-49.pt':
+    f'{PARENT_LINK}/sync/ResNetAudio-22-08-03T23-14-49.pt',  # 3s
+    'ResNetAudio-22-08-03T23-14-28.pt':
+    f'{PARENT_LINK}/sync/ResNetAudio-22-08-03T23-14-28.pt',  # 4s
+    'ResNetAudio-22-06-24T08-10-33.pt':
+    f'{PARENT_LINK}/sync/ResNetAudio-22-06-24T08-10-33.pt',  # 5s
+    'ResNetAudio-22-06-24T17-31-07.pt':
+    f'{PARENT_LINK}/sync/ResNetAudio-22-06-24T17-31-07.pt',  # 6s
+    'ResNetAudio-22-06-24T23-57-11.pt':
+    f'{PARENT_LINK}/sync/ResNetAudio-22-06-24T23-57-11.pt',  # 7s
+    'ResNetAudio-22-06-25T04-35-42.pt':
+    f'{PARENT_LINK}/sync/ResNetAudio-22-06-25T04-35-42.pt',  # 8s
+}
+
+
+def check_if_file_exists_else_download(path, fname2link=FNAME2LINK, chunk_size=1024):
+    '''Checks if file exists, if not downloads it from the link to the path'''
+    path = Path(path)
+    if not path.exists():
+        path.parent.mkdir(exist_ok=True, parents=True)
+        link = fname2link.get(path.name, None)
+        if link is None:
+            raise ValueError(f'Cant find the checkpoint file: {path}.',
+                             f'Please download it manually and ensure the path exists.')
+        with requests.get(fname2link[path.name], stream=True) as r:
+            total_size = int(r.headers.get('content-length', 0))
+            with tqdm(total=total_size, unit='B', unit_scale=True) as pbar:
+                with open(path, 'wb') as f:
+                    for data in r.iter_content(chunk_size=chunk_size):
+                        if data:
+                            f.write(data)
+                            pbar.update(chunk_size)
+
+
+def get_md5sum(path):
+    hash_md5 = md5()
+    with open(path, 'rb') as f:
+        for chunk in iter(lambda: f.read(4096 * 8), b''):
+            hash_md5.update(chunk)
+    md5sum = hash_md5.hexdigest()
+    return md5sum

selva_core/ext/synchformer/video_model_builder.py

@@ -0,0 +1,277 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# Copyright 2020 Ross Wightman
+# Modified Model definition
+
+from collections import OrderedDict
+from functools import partial
+
+import torch
+import torch.nn as nn
+from timm.layers import trunc_normal_
+
+from selva_core.ext.synchformer import vit_helper
+
+
+class VisionTransformer(nn.Module):
+    """ Vision Transformer with support for patch or hybrid CNN input stage """
+
+    def __init__(self, cfg):
+        super().__init__()
+        self.img_size = cfg.DATA.TRAIN_CROP_SIZE
+        self.patch_size = cfg.VIT.PATCH_SIZE
+        self.in_chans = cfg.VIT.CHANNELS
+        if cfg.TRAIN.DATASET == "Epickitchens":
+            self.num_classes = [97, 300]
+        else:
+            self.num_classes = cfg.MODEL.NUM_CLASSES
+        self.embed_dim = cfg.VIT.EMBED_DIM
+        self.depth = cfg.VIT.DEPTH
+        self.num_heads = cfg.VIT.NUM_HEADS
+        self.mlp_ratio = cfg.VIT.MLP_RATIO
+        self.qkv_bias = cfg.VIT.QKV_BIAS
+        self.drop_rate = cfg.VIT.DROP
+        self.drop_path_rate = cfg.VIT.DROP_PATH
+        self.head_dropout = cfg.VIT.HEAD_DROPOUT
+        self.video_input = cfg.VIT.VIDEO_INPUT
+        self.temporal_resolution = cfg.VIT.TEMPORAL_RESOLUTION
+        self.use_mlp = cfg.VIT.USE_MLP
+        self.num_features = self.embed_dim
+        norm_layer = partial(nn.LayerNorm, eps=1e-6)
+        self.attn_drop_rate = cfg.VIT.ATTN_DROPOUT
+        self.head_act = cfg.VIT.HEAD_ACT
+        self.cfg = cfg
+
+        # Patch Embedding
+        self.patch_embed = vit_helper.PatchEmbed(img_size=224,
+                                                 patch_size=self.patch_size,
+                                                 in_chans=self.in_chans,
+                                                 embed_dim=self.embed_dim)
+
+        # 3D Patch Embedding
+        self.patch_embed_3d = vit_helper.PatchEmbed3D(img_size=self.img_size,
+                                                      temporal_resolution=self.temporal_resolution,
+                                                      patch_size=self.patch_size,
+                                                      in_chans=self.in_chans,
+                                                      embed_dim=self.embed_dim,
+                                                      z_block_size=self.cfg.VIT.PATCH_SIZE_TEMP)
+        self.patch_embed_3d.proj.weight.data = torch.zeros_like(
+            self.patch_embed_3d.proj.weight.data)
+
+        # Number of patches
+        if self.video_input:
+            num_patches = self.patch_embed.num_patches * self.temporal_resolution
+        else:
+            num_patches = self.patch_embed.num_patches
+        self.num_patches = num_patches
+
+        # CLS token
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
+        trunc_normal_(self.cls_token, std=.02)
+
+        # Positional embedding
+        self.pos_embed = nn.Parameter(
+            torch.zeros(1, self.patch_embed.num_patches + 1, self.embed_dim))
+        self.pos_drop = nn.Dropout(p=cfg.VIT.POS_DROPOUT)
+        trunc_normal_(self.pos_embed, std=.02)
+
+        if self.cfg.VIT.POS_EMBED == "joint":
+            self.st_embed = nn.Parameter(torch.zeros(1, num_patches + 1, self.embed_dim))
+            trunc_normal_(self.st_embed, std=.02)
+        elif self.cfg.VIT.POS_EMBED == "separate":
+            self.temp_embed = nn.Parameter(torch.zeros(1, self.temporal_resolution, self.embed_dim))
+
+        # Layer Blocks
+        dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, self.depth)]
+        if self.cfg.VIT.ATTN_LAYER == "divided":
+            self.blocks = nn.ModuleList([
+                vit_helper.DividedSpaceTimeBlock(
+                    attn_type=cfg.VIT.ATTN_LAYER,
+                    dim=self.embed_dim,
+                    num_heads=self.num_heads,
+                    mlp_ratio=self.mlp_ratio,
+                    qkv_bias=self.qkv_bias,
+                    drop=self.drop_rate,
+                    attn_drop=self.attn_drop_rate,
+                    drop_path=dpr[i],
+                    norm_layer=norm_layer,
+                ) for i in range(self.depth)
+            ])
+        else:
+            self.blocks = nn.ModuleList([
+                vit_helper.Block(attn_type=cfg.VIT.ATTN_LAYER,
+                                 dim=self.embed_dim,
+                                 num_heads=self.num_heads,
+                                 mlp_ratio=self.mlp_ratio,
+                                 qkv_bias=self.qkv_bias,
+                                 drop=self.drop_rate,
+                                 attn_drop=self.attn_drop_rate,
+                                 drop_path=dpr[i],
+                                 norm_layer=norm_layer,
+                                 use_original_code=self.cfg.VIT.USE_ORIGINAL_TRAJ_ATTN_CODE)
+                for i in range(self.depth)
+            ])
+        self.norm = norm_layer(self.embed_dim)
+
+        # MLP head
+        if self.use_mlp:
+            hidden_dim = self.embed_dim
+            if self.head_act == 'tanh':
+                # logging.info("Using TanH activation in MLP")
+                act = nn.Tanh()
+            elif self.head_act == 'gelu':
+                # logging.info("Using GELU activation in MLP")
+                act = nn.GELU()
+            else:
+                # logging.info("Using ReLU activation in MLP")
+                act = nn.ReLU()
+            self.pre_logits = nn.Sequential(
+                OrderedDict([
+                    ('fc', nn.Linear(self.embed_dim, hidden_dim)),
+                    ('act', act),
+                ]))
+        else:
+            self.pre_logits = nn.Identity()
+
+        # Classifier Head
+        self.head_drop = nn.Dropout(p=self.head_dropout)
+        if isinstance(self.num_classes, (list, )) and len(self.num_classes) > 1:
+            for a, i in enumerate(range(len(self.num_classes))):
+                setattr(self, "head%d" % a, nn.Linear(self.embed_dim, self.num_classes[i]))
+        else:
+            self.head = nn.Linear(self.embed_dim,
+                                  self.num_classes) if self.num_classes > 0 else nn.Identity()
+
+        # Initialize weights
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        if self.cfg.VIT.POS_EMBED == "joint":
+            return {'pos_embed', 'cls_token', 'st_embed'}
+        else:
+            return {'pos_embed', 'cls_token', 'temp_embed'}
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = (nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity())
+
+    def forward_features(self, x):
+        # if self.video_input:
+        # x = x[0]
+        B = x.shape[0]
+
+        # Tokenize input
+        # if self.cfg.VIT.PATCH_SIZE_TEMP > 1:
+        # for simplicity of mapping between content dimensions (input x) and token dims (after patching)
+        # we use the same trick as for AST (see modeling_ast.ASTModel.forward for the details):
+
+        # apply patching on input
+        x = self.patch_embed_3d(x)
+        tok_mask = None
+
+        # else:
+        #     tok_mask = None
+        #     # 2D tokenization
+        #     if self.video_input:
+        #         x = x.permute(0, 2, 1, 3, 4)
+        #         (B, T, C, H, W) = x.shape
+        #         x = x.reshape(B * T, C, H, W)
+
+        #     x = self.patch_embed(x)
+
+        #     if self.video_input:
+        #         (B2, T2, D2) = x.shape
+        #         x = x.reshape(B, T * T2, D2)
+
+        # Append CLS token
+        cls_tokens = self.cls_token.expand(B, -1, -1)
+        x = torch.cat((cls_tokens, x), dim=1)
+        # if tok_mask is not None:
+        #     # prepend 1(=keep) to the mask to account for the CLS token as well
+        #     tok_mask = torch.cat((torch.ones_like(tok_mask[:, [0]]), tok_mask), dim=1)
+
+        # Interpolate positinoal embeddings
+        # if self.cfg.DATA.TRAIN_CROP_SIZE != 224:
+        #     pos_embed = self.pos_embed
+        #     N = pos_embed.shape[1] - 1
+        #     npatch = int((x.size(1) - 1) / self.temporal_resolution)
+        #     class_emb = pos_embed[:, 0]
+        #     pos_embed = pos_embed[:, 1:]
+        #     dim = x.shape[-1]
+        #     pos_embed = torch.nn.functional.interpolate(
+        #         pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2),
+        #         scale_factor=math.sqrt(npatch / N),
+        #         mode='bicubic',
+        #     )
+        #     pos_embed = pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        #     new_pos_embed = torch.cat((class_emb.unsqueeze(0), pos_embed), dim=1)
+        # else:
+        new_pos_embed = self.pos_embed
+        npatch = self.patch_embed.num_patches
+
+        # Add positional embeddings to input
+        if self.video_input:
+            if self.cfg.VIT.POS_EMBED == "separate":
+                cls_embed = self.pos_embed[:, 0, :].unsqueeze(1)
+                tile_pos_embed = new_pos_embed[:, 1:, :].repeat(1, self.temporal_resolution, 1)
+                tile_temporal_embed = self.temp_embed.repeat_interleave(npatch, 1)
+                total_pos_embed = tile_pos_embed + tile_temporal_embed
+                total_pos_embed = torch.cat([cls_embed, total_pos_embed], dim=1)
+                x = x + total_pos_embed
+            elif self.cfg.VIT.POS_EMBED == "joint":
+                x = x + self.st_embed
+        else:
+            # image input
+            x = x + new_pos_embed
+
+        # Apply positional dropout
+        x = self.pos_drop(x)
+
+        # Encoding using transformer layers
+        for i, blk in enumerate(self.blocks):
+            x = blk(x,
+                    seq_len=npatch,
+                    num_frames=self.temporal_resolution,
+                    approx=self.cfg.VIT.APPROX_ATTN_TYPE,
+                    num_landmarks=self.cfg.VIT.APPROX_ATTN_DIM,
+                    tok_mask=tok_mask)
+
+        ### v-iashin: I moved it to the forward pass
+        # x = self.norm(x)[:, 0]
+        # x = self.pre_logits(x)
+        ###
+        return x, tok_mask
+
+    # def forward(self, x):
+    #     x = self.forward_features(x)
+    #     ### v-iashin: here. This should leave the same forward output as before
+    #     x = self.norm(x)[:, 0]
+    #     x = self.pre_logits(x)
+    #     ###
+    #     x = self.head_drop(x)
+    #     if isinstance(self.num_classes, (list, )) and len(self.num_classes) > 1:
+    #         output = []
+    #         for head in range(len(self.num_classes)):
+    #             x_out = getattr(self, "head%d" % head)(x)
+    #             if not self.training:
+    #                 x_out = torch.nn.functional.softmax(x_out, dim=-1)
+    #             output.append(x_out)
+    #         return output
+    #     else:
+    #         x = self.head(x)
+    #         if not self.training:
+    #             x = torch.nn.functional.softmax(x, dim=-1)
+    #         return x

selva_core/ext/synchformer/vit_helper.py

@@ -0,0 +1,399 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# Copyright 2020 Ross Wightman
+# Modified Model definition
+"""Video models."""
+
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+from timm.layers import to_2tuple
+from torch import einsum
+from torch.nn import functional as F
+
+default_cfgs = {
+    'vit_1k':
+    'https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth',
+    'vit_1k_large':
+    'https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_224-4ee7a4dc.pth',
+}
+
+
+def qkv_attn(q, k, v, tok_mask: torch.Tensor = None):
+    sim = einsum('b i d, b j d -> b i j', q, k)
+    # apply masking if provided, tok_mask is (B*S*H, N): 1s - keep; sim is (B*S*H, H, N, N)
+    if tok_mask is not None:
+        BSH, N = tok_mask.shape
+        sim = sim.masked_fill(tok_mask.view(BSH, 1, N) == 0,
+                              float('-inf'))  # 1 - broadcasts across N
+    attn = sim.softmax(dim=-1)
+    out = einsum('b i j, b j d -> b i d', attn, v)
+    return out
+
+
+class DividedAttention(nn.Module):
+
+    def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        # init to zeros
+        self.qkv.weight.data.fill_(0)
+        self.qkv.bias.data.fill_(0)
+        self.proj.weight.data.fill_(1)
+        self.proj.bias.data.fill_(0)
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, einops_from, einops_to, tok_mask: torch.Tensor = None, **einops_dims):
+        # num of heads variable
+        h = self.num_heads
+
+        # project x to q, k, v vaalues
+        q, k, v = self.qkv(x).chunk(3, dim=-1)
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+        if tok_mask is not None:
+            # replicate token mask across heads (b, n) -> (b, h, n) -> (b*h, n) -- same as qkv but w/o d
+            assert len(tok_mask.shape) == 2
+            tok_mask = tok_mask.unsqueeze(1).expand(-1, h, -1).reshape(-1, tok_mask.shape[1])
+
+        # Scale q
+        q *= self.scale
+
+        # Take out cls_q, cls_k, cls_v
+        (cls_q, q_), (cls_k, k_), (cls_v, v_) = map(lambda t: (t[:, 0:1], t[:, 1:]), (q, k, v))
+        # the same for masking
+        if tok_mask is not None:
+            cls_mask, mask_ = tok_mask[:, 0:1], tok_mask[:, 1:]
+        else:
+            cls_mask, mask_ = None, None
+
+        # let CLS token attend to key / values of all patches across time and space
+        cls_out = qkv_attn(cls_q, k, v, tok_mask=tok_mask)
+
+        # rearrange across time or space
+        q_, k_, v_ = map(lambda t: rearrange(t, f'{einops_from} -> {einops_to}', **einops_dims),
+                         (q_, k_, v_))
+
+        # expand CLS token keys and values across time or space and concat
+        r = q_.shape[0] // cls_k.shape[0]
+        cls_k, cls_v = map(lambda t: repeat(t, 'b () d -> (b r) () d', r=r), (cls_k, cls_v))
+
+        k_ = torch.cat((cls_k, k_), dim=1)
+        v_ = torch.cat((cls_v, v_), dim=1)
+
+        # the same for masking (if provided)
+        if tok_mask is not None:
+            # since mask does not have the latent dim (d), we need to remove it from einops dims
+            mask_ = rearrange(mask_, f'{einops_from} -> {einops_to}'.replace(' d', ''),
+                              **einops_dims)
+            cls_mask = repeat(cls_mask, 'b () -> (b r) ()',
+                              r=r)  # expand cls_mask across time or space
+            mask_ = torch.cat((cls_mask, mask_), dim=1)
+
+        # attention
+        out = qkv_attn(q_, k_, v_, tok_mask=mask_)
+
+        # merge back time or space
+        out = rearrange(out, f'{einops_to} -> {einops_from}', **einops_dims)
+
+        # concat back the cls token
+        out = torch.cat((cls_out, out), dim=1)
+
+        # merge back the heads
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+
+        ## to out
+        x = self.proj(out)
+        x = self.proj_drop(x)
+        return x
+
+
+class DividedSpaceTimeBlock(nn.Module):
+
+    def __init__(self,
+                 dim=768,
+                 num_heads=12,
+                 attn_type='divided',
+                 mlp_ratio=4.,
+                 qkv_bias=False,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 act_layer=nn.GELU,
+                 norm_layer=nn.LayerNorm):
+        super().__init__()
+
+        self.einops_from_space = 'b (f n) d'
+        self.einops_to_space = '(b f) n d'
+        self.einops_from_time = 'b (f n) d'
+        self.einops_to_time = '(b n) f d'
+
+        self.norm1 = norm_layer(dim)
+
+        self.attn = DividedAttention(dim,
+                                     num_heads=num_heads,
+                                     qkv_bias=qkv_bias,
+                                     attn_drop=attn_drop,
+                                     proj_drop=drop)
+
+        self.timeattn = DividedAttention(dim,
+                                         num_heads=num_heads,
+                                         qkv_bias=qkv_bias,
+                                         attn_drop=attn_drop,
+                                         proj_drop=drop)
+
+        # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.drop_path = nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim,
+                       hidden_features=mlp_hidden_dim,
+                       act_layer=act_layer,
+                       drop=drop)
+        self.norm3 = norm_layer(dim)
+
+    def forward(self,
+                x,
+                seq_len=196,
+                num_frames=8,
+                approx='none',
+                num_landmarks=128,
+                tok_mask: torch.Tensor = None):
+        time_output = self.timeattn(self.norm3(x),
+                                    self.einops_from_time,
+                                    self.einops_to_time,
+                                    n=seq_len,
+                                    tok_mask=tok_mask)
+        time_residual = x + time_output
+
+        space_output = self.attn(self.norm1(time_residual),
+                                 self.einops_from_space,
+                                 self.einops_to_space,
+                                 f=num_frames,
+                                 tok_mask=tok_mask)
+        space_residual = time_residual + self.drop_path(space_output)
+
+        x = space_residual
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class Mlp(nn.Module):
+
+    def __init__(self,
+                 in_features,
+                 hidden_features=None,
+                 out_features=None,
+                 act_layer=nn.GELU,
+                 drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        img_size = img_size if type(img_size) is tuple else to_2tuple(img_size)
+        patch_size = img_size if type(patch_size) is tuple else to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        return x
+
+
+class PatchEmbed3D(nn.Module):
+    """ Image to Patch Embedding """
+
+    def __init__(self,
+                 img_size=224,
+                 temporal_resolution=4,
+                 in_chans=3,
+                 patch_size=16,
+                 z_block_size=2,
+                 embed_dim=768,
+                 flatten=True):
+        super().__init__()
+        self.height = (img_size // patch_size)
+        self.width = (img_size // patch_size)
+        ### v-iashin: these two are incorrect
+        # self.frames = (temporal_resolution // z_block_size)
+        # self.num_patches = self.height * self.width * self.frames
+        self.z_block_size = z_block_size
+        ###
+        self.proj = nn.Conv3d(in_chans,
+                              embed_dim,
+                              kernel_size=(z_block_size, patch_size, patch_size),
+                              stride=(z_block_size, patch_size, patch_size))
+        self.flatten = flatten
+
+    def forward(self, x):
+        B, C, T, H, W = x.shape
+        x = self.proj(x)
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)
+        return x
+
+
+class HeadMLP(nn.Module):
+
+    def __init__(self, n_input, n_classes, n_hidden=512, p=0.1):
+        super(HeadMLP, self).__init__()
+        self.n_input = n_input
+        self.n_classes = n_classes
+        self.n_hidden = n_hidden
+        if n_hidden is None:
+            # use linear classifier
+            self.block_forward = nn.Sequential(nn.Dropout(p=p),
+                                               nn.Linear(n_input, n_classes, bias=True))
+        else:
+            # use simple MLP classifier
+            self.block_forward = nn.Sequential(nn.Dropout(p=p),
+                                               nn.Linear(n_input, n_hidden, bias=True),
+                                               nn.BatchNorm1d(n_hidden), nn.ReLU(inplace=True),
+                                               nn.Dropout(p=p),
+                                               nn.Linear(n_hidden, n_classes, bias=True))
+        print(f"Dropout-NLP: {p}")
+
+    def forward(self, x):
+        return self.block_forward(x)
+
+
+def _conv_filter(state_dict, patch_size=16):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    for k, v in state_dict.items():
+        if 'patch_embed.proj.weight' in k:
+            v = v.reshape((v.shape[0], 3, patch_size, patch_size))
+        out_dict[k] = v
+    return out_dict
+
+
+def adapt_input_conv(in_chans, conv_weight, agg='sum'):
+    conv_type = conv_weight.dtype
+    conv_weight = conv_weight.float()
+    O, I, J, K = conv_weight.shape
+    if in_chans == 1:
+        if I > 3:
+            assert conv_weight.shape[1] % 3 == 0
+            # For models with space2depth stems
+            conv_weight = conv_weight.reshape(O, I // 3, 3, J, K)
+            conv_weight = conv_weight.sum(dim=2, keepdim=False)
+        else:
+            if agg == 'sum':
+                print("Summing conv1 weights")
+                conv_weight = conv_weight.sum(dim=1, keepdim=True)
+            else:
+                print("Averaging conv1 weights")
+                conv_weight = conv_weight.mean(dim=1, keepdim=True)
+    elif in_chans != 3:
+        if I != 3:
+            raise NotImplementedError('Weight format not supported by conversion.')
+        else:
+            if agg == 'sum':
+                print("Summing conv1 weights")
+                repeat = int(math.ceil(in_chans / 3))
+                conv_weight = conv_weight.repeat(1, repeat, 1, 1)[:, :in_chans, :, :]
+                conv_weight *= (3 / float(in_chans))
+            else:
+                print("Averaging conv1 weights")
+                conv_weight = conv_weight.mean(dim=1, keepdim=True)
+                conv_weight = conv_weight.repeat(1, in_chans, 1, 1)
+    conv_weight = conv_weight.to(conv_type)
+    return conv_weight
+
+
+def load_pretrained(model,
+                    cfg=None,
+                    num_classes=1000,
+                    in_chans=3,
+                    filter_fn=None,
+                    strict=True,
+                    progress=False):
+    # Load state dict
+    assert (f"{cfg.VIT.PRETRAINED_WEIGHTS} not in [vit_1k, vit_1k_large]")
+    state_dict = torch.hub.load_state_dict_from_url(url=default_cfgs[cfg.VIT.PRETRAINED_WEIGHTS])
+
+    if filter_fn is not None:
+        state_dict = filter_fn(state_dict)
+
+    input_convs = 'patch_embed.proj'
+    if input_convs is not None and in_chans != 3:
+        if isinstance(input_convs, str):
+            input_convs = (input_convs, )
+        for input_conv_name in input_convs:
+            weight_name = input_conv_name + '.weight'
+            try:
+                state_dict[weight_name] = adapt_input_conv(in_chans,
+                                                           state_dict[weight_name],
+                                                           agg='avg')
+                print(
+                    f'Converted input conv {input_conv_name} pretrained weights from 3 to {in_chans} channel(s)'
+                )
+            except NotImplementedError as e:
+                del state_dict[weight_name]
+                strict = False
+                print(
+                    f'Unable to convert pretrained {input_conv_name} weights, using random init for this layer.'
+                )
+
+    classifier_name = 'head'
+    label_offset = cfg.get('label_offset', 0)
+    pretrain_classes = 1000
+    if num_classes != pretrain_classes:
+        # completely discard fully connected if model num_classes doesn't match pretrained weights
+        del state_dict[classifier_name + '.weight']
+        del state_dict[classifier_name + '.bias']
+        strict = False
+    elif label_offset > 0:
+        # special case for pretrained weights with an extra background class in pretrained weights
+        classifier_weight = state_dict[classifier_name + '.weight']
+        state_dict[classifier_name + '.weight'] = classifier_weight[label_offset:]
+        classifier_bias = state_dict[classifier_name + '.bias']
+        state_dict[classifier_name + '.bias'] = classifier_bias[label_offset:]
+
+    loaded_state = state_dict
+    self_state = model.state_dict()
+    all_names = set(self_state.keys())
+    saved_names = set([])
+    for name, param in loaded_state.items():
+        param = param
+        if 'module.' in name:
+            name = name.replace('module.', '')
+        if name in self_state.keys() and param.shape == self_state[name].shape:
+            saved_names.add(name)
+            self_state[name].copy_(param)
+        else:
+            print(f"didnt load: {name} of shape: {param.shape}")
+    print("Missing Keys:")
+    print(all_names - saved_names)

selva_core/model/embeddings.py

@@ -0,0 +1,49 @@
+import torch
+import torch.nn as nn
+
+# https://github.com/facebookresearch/DiT
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+
+    def __init__(self, dim, frequency_embedding_size, max_period):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, dim),
+            nn.SiLU(),
+            nn.Linear(dim, dim),
+        )
+        self.dim = dim
+        self.max_period = max_period
+        assert dim % 2 == 0, 'dim must be even.'
+
+        with torch.autocast('cuda', enabled=False):
+            self.freqs = nn.Buffer(
+                1.0 / (10000**(torch.arange(0, frequency_embedding_size, 2, dtype=torch.float32) /
+                               frequency_embedding_size)),
+                persistent=False)
+            freq_scale = 10000 / max_period
+            self.freqs = freq_scale * self.freqs
+
+    def timestep_embedding(self, t):
+        """
+        Create sinusoidal timestep embeddings.
+        :param t: a 1-D Tensor of N indices, one per batch element.
+                          These may be fractional.
+        :param dim: the dimension of the output.
+        :param max_period: controls the minimum frequency of the embeddings.
+        :return: an (N, D) Tensor of positional embeddings.
+        """
+        # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+
+        args = t[:, None].float() * self.freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t).to(t.dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb

selva_core/model/flow_matching.py

@@ -0,0 +1,87 @@
+import logging
+from typing import Callable, Optional
+
+import torch
+from torchdiffeq import odeint
+
+log = logging.getLogger()
+
+
+# Partially from https://github.com/gle-bellier/flow-matching
+class FlowMatching:
+
+    def __init__(self, min_sigma: float = 0.0, inference_mode='euler', num_steps: int = 25,
+                 target: str = 'v'):
+        # inference_mode: 'euler' or 'adaptive'
+        # num_steps: number of steps in the euler inference mode
+        super().__init__()
+        self.min_sigma = min_sigma
+        self.inference_mode = inference_mode
+        self.num_steps = num_steps
+        self.target = target
+
+        # self.fm = ExactOptimalTransportConditionalFlowMatcher(sigma=min_sigma)
+
+        assert self.inference_mode in ['euler', 'adaptive']
+        if self.inference_mode == 'adaptive' and num_steps > 0:
+            log.info('The number of steps is ignored in adaptive inference mode ')
+
+    def get_conditional_flow(self, x0: torch.Tensor, x1: torch.Tensor,
+                             t: torch.Tensor) -> torch.Tensor:
+        # which is psi_t(x), eq 22 in flow matching for generative models
+        t = t[:, None, None].expand_as(x0)
+        return (1 - (1 - self.min_sigma) * t) * x0 + t * x1
+
+    def loss(self, predicted_v: torch.Tensor, x0: torch.Tensor, x1: torch.Tensor,
+             xt: Optional[torch.Tensor] = None, t: Optional[torch.Tensor] = None) -> torch.Tensor:
+        # return the mean error without reducing the batch dimension
+        reduce_dim = list(range(1, len(predicted_v.shape)))
+        if self.target == 'v':
+            target_v = x1 - (1 - self.min_sigma) * x0
+            return (predicted_v - target_v).pow(2).mean(dim=reduce_dim)
+        elif self.target == 'x1':
+            if xt is None or t is None:
+                raise ValueError("xt and t must be provided when target is 'x1'")
+            t = t[:, None, None].expand_as(x0)
+            predicted_x1 = xt + (1 - t) * predicted_v - self.min_sigma * x0
+            return (predicted_x1 - x1).pow(2).mean(dim=reduce_dim)
+        else:
+            raise ValueError(f"Unknown target: {self.target}. Supported targets are 'v' and 'x1'.")
+
+    def get_x0_xt_c(
+        self,
+        x1: torch.Tensor,
+        t: torch.Tensor,
+        Cs: list[torch.Tensor],
+        generator: Optional[torch.Generator] = None
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        x0 = torch.empty_like(x1).normal_(generator=generator)
+
+        xt = self.get_conditional_flow(x0, x1, t)
+        return x0, x1, xt, Cs
+
+    def to_prior(self, fn: Callable, x1: torch.Tensor) -> torch.Tensor:
+        return self.run_t0_to_t1(fn, x1, 1, 0)
+
+    def to_data(self, fn: Callable, x0: torch.Tensor) -> torch.Tensor:
+        return self.run_t0_to_t1(fn, x0, 0, 1)
+
+    def run_t0_to_t1(self, fn: Callable, x0: torch.Tensor, t0: float, t1: float) -> torch.Tensor:
+        # fn: a function that takes (t, x) and returns the direction x0->x1
+
+        if self.inference_mode == 'adaptive':
+            return odeint(fn, x0, torch.tensor([t0, t1], device=x0.device, dtype=x0.dtype))
+        elif self.inference_mode == 'euler':
+            x = x0
+            steps = torch.linspace(t0, t1 - self.min_sigma, self.num_steps + 1)
+            for ti, t in enumerate(steps[:-1]):
+                flow = fn(t, x)
+                next_t = steps[ti + 1]
+                dt = next_t - t
+                x = x + dt * flow
+                # print(f"DEBUG timestep {ti=}")
+                # if ti == 11:
+                #     print(f'{ti=} quit!!!!!!!!!!!!')
+                #     quit();
+
+        return x

selva_core/model/low_level.py

@@ -0,0 +1,95 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+class ChannelLastConv1d(nn.Conv1d):
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.permute(0, 2, 1)
+        x = super().forward(x)
+        x = x.permute(0, 2, 1)
+        return x
+
+
+# https://github.com/Stability-AI/sd3-ref
+class MLP(nn.Module):
+
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int = 256,
+    ):
+        """
+        Initialize the FeedForward module.
+
+        Args:
+            dim (int): Input dimension.
+            hidden_dim (int): Hidden dimension of the feedforward layer.
+            multiple_of (int): Value to ensure hidden dimension is a multiple of this value.
+
+        Attributes:
+            w1 (ColumnParallelLinear): Linear transformation for the first layer.
+            w2 (RowParallelLinear): Linear transformation for the second layer.
+            w3 (ColumnParallelLinear): Linear transformation for the third layer.
+
+        """
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = nn.Linear(dim, hidden_dim, bias=False)
+        self.w2 = nn.Linear(hidden_dim, dim, bias=False)
+        self.w3 = nn.Linear(dim, hidden_dim, bias=False)
+
+    def forward(self, x):
+        return self.w2(F.silu(self.w1(x)) * self.w3(x))
+
+
+class ConvMLP(nn.Module):
+
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int = 256,
+        kernel_size: int = 3,
+        padding: int = 1,
+    ):
+        """
+        Initialize the FeedForward module.
+
+        Args:
+            dim (int): Input dimension.
+            hidden_dim (int): Hidden dimension of the feedforward layer.
+            multiple_of (int): Value to ensure hidden dimension is a multiple of this value.
+
+        Attributes:
+            w1 (ColumnParallelLinear): Linear transformation for the first layer.
+            w2 (RowParallelLinear): Linear transformation for the second layer.
+            w3 (ColumnParallelLinear): Linear transformation for the third layer.
+
+        """
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = ChannelLastConv1d(dim,
+                                    hidden_dim,
+                                    bias=False,
+                                    kernel_size=kernel_size,
+                                    padding=padding)
+        self.w2 = ChannelLastConv1d(hidden_dim,
+                                    dim,
+                                    bias=False,
+                                    kernel_size=kernel_size,
+                                    padding=padding)
+        self.w3 = ChannelLastConv1d(dim,
+                                    hidden_dim,
+                                    bias=False,
+                                    kernel_size=kernel_size,
+                                    padding=padding)
+
+    def forward(self, x):
+        return self.w2(F.silu(self.w1(x)) * self.w3(x))

selva_core/model/networks_generator.py

@@ -0,0 +1,475 @@
+import logging
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from selva_core.ext.rotary_embeddings import compute_rope_rotations
+from selva_core.model.embeddings import TimestepEmbedder
+from selva_core.model.low_level import MLP, ChannelLastConv1d, ConvMLP
+from selva_core.model.transformer_layers import (FinalBlock, JointBlock, MMDitSingleBlock)
+
+log = logging.getLogger()
+
+
+@dataclass
+class PreprocessedConditions:
+    clip_f: torch.Tensor
+    sync_f: torch.Tensor
+    text_f: torch.Tensor
+    clip_f_c: torch.Tensor
+    text_f_c: torch.Tensor
+
+
+# Partially from https://github.com/facebookresearch/DiT
+class MMAudio(nn.Module):
+
+    def __init__(self,
+                 *,
+                 latent_dim: int,
+                 clip_dim: int,
+                 sync_dim: int,
+                 text_dim: int,
+                 hidden_dim: int,
+                 depth: int,
+                 fused_depth: int,
+                 num_heads: int,
+                 mlp_ratio: float = 4.0,
+                 latent_seq_len: int,
+                 clip_seq_len: int,
+                 sync_seq_len: int,
+                 text_seq_len: int = 77,
+                 latent_mean: Optional[torch.Tensor] = None,
+                 latent_std: Optional[torch.Tensor] = None,
+                 empty_string_feat: Optional[torch.Tensor] = None,
+                 v2: bool = False) -> None:
+        super().__init__()
+
+        self.v2 = v2
+        self.latent_dim = latent_dim
+        self._latent_seq_len = latent_seq_len
+        self._clip_seq_len = clip_seq_len
+        self._sync_seq_len = sync_seq_len
+        self._text_seq_len = text_seq_len
+        self.hidden_dim = hidden_dim
+        self.num_heads = num_heads
+
+        if v2:
+            self.audio_input_proj = nn.Sequential(
+                ChannelLastConv1d(latent_dim, hidden_dim, kernel_size=7, padding=3),
+                nn.SiLU(),
+                ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=7, padding=3),
+            )
+
+            self.clip_input_proj = nn.Sequential(
+                nn.Linear(clip_dim, hidden_dim),
+                nn.SiLU(),
+                ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1),
+            )
+
+            self.sync_input_proj = nn.Sequential(
+                ChannelLastConv1d(sync_dim, hidden_dim, kernel_size=7, padding=3),
+                nn.SiLU(),
+                ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1),
+            )
+
+            self.text_input_proj = nn.Sequential(
+                nn.Linear(text_dim, hidden_dim),
+                nn.SiLU(),
+                MLP(hidden_dim, hidden_dim * 4),
+            )
+        else:
+            self.audio_input_proj = nn.Sequential(
+                ChannelLastConv1d(latent_dim, hidden_dim, kernel_size=7, padding=3),
+                nn.SELU(),
+                ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=7, padding=3),
+            )
+
+            self.clip_input_proj = nn.Sequential(
+                nn.Linear(clip_dim, hidden_dim),
+                ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1),
+            )
+
+            self.sync_input_proj = nn.Sequential(
+                ChannelLastConv1d(sync_dim, hidden_dim, kernel_size=7, padding=3),
+                nn.SELU(),
+                ConvMLP(hidden_dim, hidden_dim * 4, kernel_size=3, padding=1),
+            )
+
+            self.text_input_proj = nn.Sequential(
+                nn.Linear(text_dim, hidden_dim),
+                MLP(hidden_dim, hidden_dim * 4),
+            )
+
+        self.clip_cond_proj = nn.Linear(hidden_dim, hidden_dim)
+        self.text_cond_proj = nn.Linear(hidden_dim, hidden_dim)
+        self.global_cond_mlp = MLP(hidden_dim, hidden_dim * 4)
+        # each synchformer output segment has 8 feature frames
+        self.sync_pos_emb = nn.Parameter(torch.zeros((1, 1, 8, sync_dim)))
+
+        self.final_layer = FinalBlock(hidden_dim, latent_dim)
+
+        if v2:
+            self.t_embed = TimestepEmbedder(hidden_dim,
+                                            frequency_embedding_size=hidden_dim,
+                                            max_period=1)
+        else:
+            self.t_embed = TimestepEmbedder(hidden_dim,
+                                            frequency_embedding_size=256,
+                                            max_period=10000)
+        self.joint_blocks = nn.ModuleList([
+            JointBlock(hidden_dim,
+                       num_heads,
+                       mlp_ratio=mlp_ratio,
+                       pre_only=(i == depth - fused_depth - 1)) for i in range(depth - fused_depth)
+        ])
+
+        self.fused_blocks = nn.ModuleList([
+            MMDitSingleBlock(hidden_dim, num_heads, mlp_ratio=mlp_ratio, kernel_size=3, padding=1)
+            for i in range(fused_depth)
+        ])
+
+        if latent_mean is None:
+            # these values are not meant to be used
+            # if you don't provide mean/std here, we should load them later from a checkpoint
+            assert latent_std is None
+            latent_mean = torch.ones(latent_dim).view(1, 1, -1).fill_(float('nan'))
+            latent_std = torch.ones(latent_dim).view(1, 1, -1).fill_(float('nan'))
+        else:
+            assert latent_std is not None
+            assert latent_mean.numel() == latent_dim, f'{latent_mean.numel()=} != {latent_dim=}'
+        if empty_string_feat is None:
+            empty_string_feat = torch.zeros((text_seq_len, text_dim))
+        self.latent_mean = nn.Parameter(latent_mean.view(1, 1, -1), requires_grad=False)
+        self.latent_std = nn.Parameter(latent_std.view(1, 1, -1), requires_grad=False)
+
+        self.empty_string_feat = nn.Parameter(empty_string_feat, requires_grad=False)
+        self.empty_clip_feat = nn.Parameter(torch.zeros(1, clip_dim), requires_grad=True)
+        self.empty_sync_feat = nn.Parameter(torch.zeros(1, sync_dim), requires_grad=True)
+
+        self.initialize_weights()
+        self.initialize_rotations()
+
+    def initialize_rotations(self):
+        base_freq = 1.0
+        latent_rot = compute_rope_rotations(self._latent_seq_len,
+                                            self.hidden_dim // self.num_heads,
+                                            10000,
+                                            freq_scaling=base_freq,
+                                            device=self.device)
+        clip_rot = compute_rope_rotations(self._clip_seq_len,
+                                          self.hidden_dim // self.num_heads,
+                                          10000,
+                                          freq_scaling=base_freq * self._latent_seq_len /
+                                          self._clip_seq_len,
+                                          device=self.device)
+
+        self.latent_rot = nn.Buffer(latent_rot, persistent=False)
+        self.clip_rot = nn.Buffer(clip_rot, persistent=False)
+
+    def update_seq_lengths(self, latent_seq_len: int, clip_seq_len: int, sync_seq_len: int) -> None:
+        self._latent_seq_len = latent_seq_len
+        self._clip_seq_len = clip_seq_len
+        self._sync_seq_len = sync_seq_len
+        self.initialize_rotations()
+
+    def initialize_weights(self):
+
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+
+        self.apply(_basic_init)
+
+        # Initialize timestep embedding MLP:
+        nn.init.normal_(self.t_embed.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.t_embed.mlp[2].weight, std=0.02)
+
+        # Zero-out adaLN modulation layers in DiT blocks:
+        for block in self.joint_blocks:
+            nn.init.constant_(block.latent_block.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(block.latent_block.adaLN_modulation[-1].bias, 0)
+            nn.init.constant_(block.clip_block.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(block.clip_block.adaLN_modulation[-1].bias, 0)
+            nn.init.constant_(block.text_block.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(block.text_block.adaLN_modulation[-1].bias, 0)
+        for block in self.fused_blocks:
+            nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+        # Zero-out output layers:
+        nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
+        nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
+        nn.init.constant_(self.final_layer.conv.weight, 0)
+        nn.init.constant_(self.final_layer.conv.bias, 0)
+
+        # empty string feat shall be initialized by a CLIP encoder
+        nn.init.constant_(self.sync_pos_emb, 0)
+        nn.init.constant_(self.empty_clip_feat, 0)
+        nn.init.constant_(self.empty_sync_feat, 0)
+
+    def normalize(self, x: torch.Tensor) -> torch.Tensor:
+        # return (x - self.latent_mean) / self.latent_std
+        return x.sub_(self.latent_mean).div_(self.latent_std)
+
+    def unnormalize(self, x: torch.Tensor) -> torch.Tensor:
+        # return x * self.latent_std + self.latent_mean
+        return x.mul_(self.latent_std).add_(self.latent_mean)
+
+    def preprocess_conditions(self, clip_f: torch.Tensor, sync_f: torch.Tensor,
+                              text_f: torch.Tensor) -> PreprocessedConditions:
+        """
+        cache computations that do not depend on the latent/time step
+        i.e., the features are reused over steps during inference
+        """
+        assert clip_f.shape[1] == self._clip_seq_len, f'{clip_f.shape=} {self._clip_seq_len=}'
+        assert sync_f.shape[1] == self._sync_seq_len, f'{sync_f.shape=} {self._sync_seq_len=}'
+        assert text_f.shape[1] == self._text_seq_len, f'{text_f.shape=} {self._text_seq_len=}'
+
+        bs = clip_f.shape[0]
+
+        # B * num_segments (24) * 8 * 768
+        num_sync_segments = self._sync_seq_len // 8
+        sync_f = sync_f.view(bs, num_sync_segments, 8, -1) + self.sync_pos_emb
+        sync_f = sync_f.flatten(1, 2)  # (B, VN, D)
+
+        # extend vf to match x
+        clip_f = self.clip_input_proj(clip_f)  # (B, VN, D)
+        sync_f = self.sync_input_proj(sync_f)  # (B, VN, D)
+        text_f = self.text_input_proj(text_f)  # (B, VN, D)
+
+        # upsample the sync features to match the audio
+        sync_f = sync_f.transpose(1, 2)  # (B, D, VN)
+        sync_f = F.interpolate(sync_f, size=self._latent_seq_len, mode='nearest-exact')
+        sync_f = sync_f.transpose(1, 2)  # (B, N, D)
+
+        # get conditional features from the clip side
+        clip_f_c = self.clip_cond_proj(clip_f.mean(dim=1))  # (B, D)
+        text_f_c = self.text_cond_proj(text_f.mean(dim=1))  # (B, D)
+
+        return PreprocessedConditions(clip_f=clip_f,
+                                      sync_f=sync_f,
+                                      text_f=text_f,
+                                      clip_f_c=clip_f_c,
+                                      text_f_c=text_f_c)
+
+    def predict_flow(self, latent: torch.Tensor, t: torch.Tensor,
+                     conditions: PreprocessedConditions) -> torch.Tensor:
+        """
+        for non-cacheable computations
+        """
+        assert latent.shape[1] == self._latent_seq_len, f'{latent.shape=} {self._latent_seq_len=}'
+
+        clip_f = conditions.clip_f
+        sync_f = conditions.sync_f
+        text_f = conditions.text_f
+        clip_f_c = conditions.clip_f_c
+        text_f_c = conditions.text_f_c
+
+        latent = self.audio_input_proj(latent)  # (B, N, D)
+        global_c = self.global_cond_mlp(clip_f_c + text_f_c)  # (B, D)
+
+        global_c = self.t_embed(t).unsqueeze(1) + global_c.unsqueeze(1)  # (B, D)
+        extended_c = global_c + sync_f
+
+        for block in self.joint_blocks:
+        # for i, block in enumerate(self.joint_blocks):
+        #     # debug attention weight attn map
+        #     block.forward_debug(latent.clone(), clip_f.clone(), text_f.clone(), 
+        #                         global_c.clone(), extended_c.clone(),
+        #                         self.latent_rot, self.clip_rot,
+        #                         layer_idx=i+1)
+            latent, clip_f, text_f = block(latent, clip_f, text_f, global_c, extended_c,
+                                           self.latent_rot, self.clip_rot)  # (B, N, D)
+
+        for block in self.fused_blocks:
+            latent = block(latent, extended_c, self.latent_rot)
+
+        flow = self.final_layer(latent, global_c)  # (B, N, out_dim), remove t
+        return flow
+
+    def forward(self, latent: torch.Tensor, clip_f: torch.Tensor, sync_f: torch.Tensor,
+                text_f: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        """
+        latent: (B, N, C) 
+        vf: (B, T, C_V)
+        t: (B,)
+        """
+        conditions = self.preprocess_conditions(clip_f, sync_f, text_f)
+        flow = self.predict_flow(latent, t, conditions)
+        return flow
+
+    def get_empty_string_sequence(self, bs: int) -> torch.Tensor:
+        return self.empty_string_feat.unsqueeze(0).expand(bs, -1, -1)
+
+    def get_empty_clip_sequence(self, bs: int) -> torch.Tensor:
+        return self.empty_clip_feat.unsqueeze(0).expand(bs, self._clip_seq_len, -1)
+
+    def get_empty_sync_sequence(self, bs: int) -> torch.Tensor:
+        return self.empty_sync_feat.unsqueeze(0).expand(bs, self._sync_seq_len, -1)
+
+    def get_empty_conditions(
+            self,
+            bs: int,
+            *,
+            negative_text_features: Optional[torch.Tensor] = None) -> PreprocessedConditions:
+        if negative_text_features is not None:
+            empty_text = negative_text_features
+        else:
+            empty_text = self.get_empty_string_sequence(1)
+
+        empty_clip = self.get_empty_clip_sequence(1)
+        empty_sync = self.get_empty_sync_sequence(1)
+        conditions = self.preprocess_conditions(empty_clip, empty_sync, empty_text)
+        conditions.clip_f = conditions.clip_f.expand(bs, -1, -1)
+        conditions.sync_f = conditions.sync_f.expand(bs, -1, -1)
+        conditions.clip_f_c = conditions.clip_f_c.expand(bs, -1)
+        if negative_text_features is None:
+            conditions.text_f = conditions.text_f.expand(bs, -1, -1)
+            conditions.text_f_c = conditions.text_f_c.expand(bs, -1)
+
+        return conditions
+
+    def ode_wrapper(self, t: torch.Tensor, latent: torch.Tensor, conditions: PreprocessedConditions,
+                    empty_conditions: PreprocessedConditions, cfg_strength: float) -> torch.Tensor:
+        t = t * torch.ones(len(latent), device=latent.device, dtype=latent.dtype)
+
+        if cfg_strength < 1.0:
+            return self.predict_flow(latent, t, conditions)
+        else:
+            return (cfg_strength * self.predict_flow(latent, t, conditions) +
+                    (1 - cfg_strength) * self.predict_flow(latent, t, empty_conditions))
+
+    def load_weights(self, src_dict) -> None:
+        if 't_embed.freqs' in src_dict:
+            del src_dict['t_embed.freqs']
+        if 'latent_rot' in src_dict:
+            del src_dict['latent_rot']
+        if 'clip_rot' in src_dict:
+            del src_dict['clip_rot']
+
+        self.load_state_dict(src_dict, strict=True)
+
+    @property
+    def device(self) -> torch.device:
+        return self.latent_mean.device
+
+    @property
+    def latent_seq_len(self) -> int:
+        return self._latent_seq_len
+
+    @property
+    def clip_seq_len(self) -> int:
+        return self._clip_seq_len
+
+    @property
+    def sync_seq_len(self) -> int:
+        return self._sync_seq_len
+
+
+def small_16k(**kwargs) -> MMAudio:
+    num_heads = 7
+    return MMAudio(latent_dim=20,
+                   clip_dim=1024,
+                   sync_dim=768,
+                   text_dim=1024,
+                   hidden_dim=64 * num_heads,
+                   depth=12,
+                   fused_depth=8,
+                   num_heads=num_heads,
+                   latent_seq_len=250,
+                   clip_seq_len=64,
+                   sync_seq_len=192,
+                   **kwargs)
+
+
+def small_44k(**kwargs) -> MMAudio:
+    num_heads = 7
+    return MMAudio(latent_dim=40,
+                   clip_dim=1024,
+                   sync_dim=768,
+                   text_dim=1024,
+                   hidden_dim=64 * num_heads,
+                   depth=12,
+                   fused_depth=8,
+                   num_heads=num_heads,
+                   latent_seq_len=345,
+                   clip_seq_len=64,
+                   sync_seq_len=192,
+                   **kwargs)
+
+
+def medium_44k(**kwargs) -> MMAudio:
+    num_heads = 14
+    return MMAudio(latent_dim=40,
+                   clip_dim=1024,
+                   sync_dim=768,
+                   text_dim=1024,
+                   hidden_dim=64 * num_heads,
+                   depth=12,
+                   fused_depth=8,
+                   num_heads=num_heads,
+                   latent_seq_len=345,
+                   clip_seq_len=64,
+                   sync_seq_len=192,
+                   **kwargs)
+
+
+def large_44k(**kwargs) -> MMAudio:
+    num_heads = 14
+    return MMAudio(latent_dim=40,
+                   clip_dim=1024,
+                   sync_dim=768,
+                   text_dim=1024,
+                   hidden_dim=64 * num_heads,
+                   depth=21,
+                   fused_depth=14,
+                   num_heads=num_heads,
+                   latent_seq_len=345,
+                   clip_seq_len=64,
+                   sync_seq_len=192,
+                   **kwargs)
+
+
+def large_44k_v2(**kwargs) -> MMAudio:
+    num_heads = 14
+    return MMAudio(latent_dim=40,
+                   clip_dim=1024,
+                   sync_dim=768,
+                   text_dim=1024,
+                   hidden_dim=64 * num_heads,
+                   depth=21,
+                   fused_depth=14,
+                   num_heads=num_heads,
+                   latent_seq_len=345,
+                   clip_seq_len=64,
+                   sync_seq_len=192,
+                   v2=True,
+                   **kwargs)
+
+
+def get_my_mmaudio(name: str, **kwargs) -> MMAudio:
+    if name == 'small_16k':
+        return small_16k(**kwargs)
+    if name == 'small_44k':
+        return small_44k(**kwargs)
+    if name == 'medium_44k':
+        return medium_44k(**kwargs)
+    if name == 'large_44k':
+        return large_44k(**kwargs)
+    if name == 'large_44k_v2':
+        return large_44k_v2(**kwargs)
+
+    raise ValueError(f'Unknown model name: {name}')
+
+
+if __name__ == '__main__':
+    network = get_my_mmaudio('small_16k')
+
+    # print the number of parameters in terms of millions
+    num_params = sum(p.numel() for p in network.parameters()) / 1e6
+    print(f'Number of parameters: {num_params:.2f}M')

selva_core/model/networks_video_enc.py

@@ -0,0 +1,188 @@
+from typing import Optional, Union, List, Tuple, Any, Mapping
+from dataclasses import dataclass
+
+import einops
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from selva_core.model.text_synchformer import TextSynchformer
+from selva_core.utils.transforms import generate_multiple_segments
+
+
+@dataclass
+class PreprocessedConditions:
+    sync_f: torch.Tensor
+    sync_f_c: torch.Tensor
+    text_f: torch.Tensor
+    text_f_c: torch.Tensor
+    text_mask: torch.Tensor
+
+
+class TextSynch(TextSynchformer):
+
+    def __init__(self,
+                 *,
+                 text_dim: int,
+                 video_seq_len: int = 192, 
+                 max_text_seq_len: int = 512,
+                 empty_string_feat: torch.Tensor = None,
+                 num_sup_text_tokens: int = 5,
+                 sync_batch_size_multiplier: Union[int, float] = -1,
+                 xattn_depth: int = 1,
+                 ) -> None:
+        super().__init__(
+            text_dim=text_dim, 
+            max_text_seq_len=max_text_seq_len,
+            xattn_depth=xattn_depth,
+        )
+
+        self._video_seq_len = video_seq_len
+        self.num_sup_text_tokens = num_sup_text_tokens
+        self.sync_batch_size_multiplier = sync_batch_size_multiplier
+        
+        if num_sup_text_tokens > 0:
+            self.sup_text_feat = nn.Parameter(torch.zeros(num_sup_text_tokens, self.text_dim),
+                                                  requires_grad=True)
+        if empty_string_feat is None:
+            empty_string_feat = torch.zeros((1, text_dim))
+        self.empty_string_feat = nn.Parameter(empty_string_feat, requires_grad=False)
+        
+        self.initialize_weights()
+    
+    def update_seq_lengths(self, video_seq_len: int) -> None:
+        self._video_seq_len = video_seq_len
+    
+    def get_empty_string_sequence(self, bs: int) -> torch.Tensor:
+        return self.empty_string_feat.unsqueeze(0).expand(bs, -1, -1)
+    
+    def get_sup_text_sequence(self, bs: int) -> torch.Tensor:
+        if self.num_sup_text_tokens <= 0:
+            raise ValueError(f'supplementary text tokens not enabled as {self.num_sup_text_tokens=}')
+        return self.sup_text_feat.expand(bs, -1, -1)
+    
+    def prepend_sup_text_tokens(self, text_f: torch.Tensor, text_mask: torch.Tensor) \
+        -> Tuple[torch.Tensor, torch.Tensor]:
+        if self.num_sup_text_tokens <= 0:
+            return text_f, text_mask
+        bs = text_f.shape[0]
+        sup_text_f = self.get_sup_text_sequence(bs) # (B, S, D)
+        sup_text_mask = torch.ones(bs, sup_text_f.shape[1], 
+                                       device=text_mask.device, dtype=text_mask.dtype) # (B, S)
+        text_f = torch.cat([sup_text_f, text_f], dim=1)
+        text_mask = torch.cat([sup_text_mask, text_mask], dim=1)
+        return text_f, text_mask
+
+    def encode_video_with_sync(self, x: torch.Tensor, text_f: torch.Tensor,
+                               text_mask: torch.Tensor) -> torch.Tensor:
+        # x: (B, T, C, H, W) H/W: 384
+
+        b, t, c, h, w = x.shape
+        assert c == 3 and h == 224 and w == 224
+
+        # partition the video
+        segment_size = 16
+        step_size = 8
+        x = generate_multiple_segments(x, segment_size, step_size) # (B, S, T, C, H, W)
+        num_segments = x.shape[1]
+        
+        outputs = []
+        if self.sync_batch_size_multiplier <= 0:
+            batch_size = b
+        else:
+            batch_size = int(b * self.sync_batch_size_multiplier)
+        x = einops.rearrange(x, 'b s t c h w -> (b s) 1 t c h w')
+        for i in range(0, b * num_segments, batch_size):
+            start_idx = i // num_segments
+            end_idx = min((i + batch_size - 1) // num_segments + 1, b)
+            text_f_batch = text_f[start_idx:end_idx]
+            text_mask_batch = text_mask[start_idx:end_idx]
+            current_total_batch_size = min(batch_size, b * num_segments - i)
+            
+            repeats = torch.zeros(end_idx - start_idx, dtype=torch.long, device=x.device)
+            for j in range(current_total_batch_size):
+                original_batch_idx = (i + j) // num_segments
+                repeats[original_batch_idx - start_idx] += 1
+
+            text_f_batch_repeated = torch.repeat_interleave(text_f_batch, repeats, dim=0)
+            text_mask_batch_repeated = torch.repeat_interleave(text_mask_batch, repeats, dim=0)
+
+            outputs.append(self.forward_vfeat(
+                x[i:i + batch_size], 
+                text_f=text_f_batch_repeated, 
+                text_mask=text_mask_batch_repeated
+            ))
+        x = torch.cat(outputs, dim=0)
+        x = einops.rearrange(x, '(b s) 1 t d -> b (s t) d', b=b)
+        return x
+    
+    def encode_audio_with_sync(self, x: torch.Tensor, text_f: torch.Tensor,
+                               text_mask: torch.Tensor) -> torch.Tensor:
+        return self.forward_afeat(
+            x, text_f=text_f, text_mask=text_mask
+        )
+    
+    def load_synchformer_state_dict(self, src_dict: dict):
+        self.load_state_dict(src_dict, strict=True)
+    
+    def load_weights(self, src_dict) -> None:
+        self.load_state_dict(src_dict, strict=True)
+
+    @property
+    def device(self) -> torch.device:
+        return self.empty_string_feat.device
+    
+    @property
+    def dtype(self) -> torch.dtype:
+        return self.empty_string_feat.dtype
+
+    @property
+    def video_seq_len(self) -> int:
+        return self._video_seq_len
+
+    @property
+    def audio_seq_len(self) -> int:
+        return self._audio_seq_len
+    
+    def load_state_dict(self, sd: Mapping[str, Any], strict: bool = True):
+        target_keys = (['vfeat_extractor'] if self.video else []) \
+                    + (['afeat_extractor'] if self.audio else []) \
+                    + ['text_proj', 'synch_text_cross_blocks', 
+                       'sup_text_feat', 'empty_string_feat']
+        # discard all entries except vfeat_extractor / afeat_extractor
+        sd = {k: v for k, v in sd.items() if any(k.startswith(tk) 
+                                                 for tk in target_keys)}
+        
+
+        return nn.Module.load_state_dict(self, sd, strict=strict)
+
+
+def depth1(**kwargs) -> TextSynch:
+    return TextSynch(text_dim=768,
+                     video_seq_len=192,
+                     max_text_seq_len=512,
+                     xattn_depth=1,
+                     **kwargs)
+    
+
+def get_my_textsynch(name: str, **kwargs) -> TextSynch:
+    if name.startswith('depth1'):
+        return depth1(**kwargs)
+    else:
+        raise ValueError(f'Unknown model name: {name}')
+
+
+if __name__ == '__main__':
+    network = get_my_textsynch('depth1')
+    # print the number of parameters in terms of millions
+    num_params = sum(p.numel() for p in network.parameters()) / 1e6
+    print(f'Number of parameters: {num_params:.2f}M')
+    
+    torch.compile(network.encode_video_with_sync)
+    print(f"Compiled encode_video_with_sync")
+    torch.compile(network.predict_flow)
+    print(f"Compiled predict_flow")
+    torch.compile(network.preprocess_conditions)
+    print(f"Compiled preprocess_conditions:")
+    torch.compile(network.forward)
+    print(f"Compiled forward:")

selva_core/model/sequence_config.py

@@ -0,0 +1,62 @@
+import dataclasses
+import math
+
+
+@dataclasses.dataclass
+class SequenceConfig:
+    # general
+    duration: float
+
+    # audio
+    sampling_rate: int
+    spectrogram_frame_rate: int
+    latent_downsample_rate: int = 2
+
+    # visual
+    clip_frame_rate: int = 8
+    sync_frame_rate: int = 25
+    sync_num_frames_per_segment: int = 16
+    sync_step_size: int = 8
+    sync_downsample_rate: int = 2
+
+    @property
+    def num_audio_frames(self) -> int:
+        # we need an integer number of latents
+        return self.latent_seq_len * self.spectrogram_frame_rate * self.latent_downsample_rate
+
+    @property
+    def latent_seq_len(self) -> int:
+        return int(
+            math.ceil(self.duration * self.sampling_rate / self.spectrogram_frame_rate /
+                      self.latent_downsample_rate))
+
+    @property
+    def clip_seq_len(self) -> int:
+        return int(self.duration * self.clip_frame_rate)
+
+    @property
+    def sync_seg_len(self) -> int:
+        num_frames = self.duration * self.sync_frame_rate
+        num_segments = (num_frames - self.sync_num_frames_per_segment) // self.sync_step_size + 1
+        return int(num_segments)
+    
+    @property
+    def sync_seq_len(self) -> int:
+        return int(self.sync_seg_len * self.sync_num_frames_per_segment / self.sync_downsample_rate)
+
+
+CONFIG_16K = SequenceConfig(duration=8.0, sampling_rate=16000, spectrogram_frame_rate=256)
+CONFIG_44K = SequenceConfig(duration=8.0, sampling_rate=44100, spectrogram_frame_rate=512)
+
+if __name__ == '__main__':
+    assert CONFIG_16K.latent_seq_len == 250
+    assert CONFIG_16K.clip_seq_len == 64
+    assert CONFIG_16K.sync_seq_len == 192
+    assert CONFIG_16K.num_audio_frames == 128000
+
+    assert CONFIG_44K.latent_seq_len == 345
+    assert CONFIG_44K.clip_seq_len == 64
+    assert CONFIG_44K.sync_seq_len == 192
+    assert CONFIG_44K.num_audio_frames == 353280
+
+    print('Passed')

selva_core/model/text_synchformer.py

@@ -0,0 +1,199 @@
+import logging
+from typing import Any, Mapping
+
+import einops
+import torch
+from torch import nn
+
+from selva_core.ext.synchformer.motionformer import MotionFormer, SpatialTransformerEncoderLayer, BaseEncoderLayer
+from selva_core.ext.synchformer.astransformer import AST
+from selva_core.model.transformer_layers import (MMCrossAttentionBlock)
+from selva_core.model.low_level import MLP
+
+
+class ExtendedMotionFormer(MotionFormer):
+    """Extended MotionFormer with additional methods for text synchronization."""
+    
+    def forward_segments_without_aggregation(self, x, orig_shape: tuple) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Extract features without spatial-temporal aggregation.
+        Args:
+            x: Input tensor of shape (BS, C, T, H, W) where S is the number of segments
+            orig_shape: Original shape tuple (B, S, C, T, H, W)
+        Returns:
+            Tuple of (features, mask) where features are of shape (B*S, D, t, h, w)
+        """
+        x, x_mask = self.forward_features(x)
+        
+        assert self.extract_features
+
+        # (BS, T, D) where T = 1 + (224 // 16) * (224 // 16) * 8
+        x = x[:, 1:, :]  # without the CLS token for efficiency
+        x = self.norm(x)
+        x = self.pre_logits(x)
+        
+        if self.factorize_space_time:
+            x = self.restore_spatio_temp_dims(x, orig_shape)  # (B*S, D, t, h, w) <- (B*S, t*h*w, D)
+        return x, x_mask
+    
+    def spatiotemporal_aggregation(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+        """
+        Apply spatial-temporal aggregation to features.
+        Args:
+            x: Features tensor of shape (B*S, D, t, h, w)
+            x_mask: Mask tensor
+        Returns:
+            Aggregated features of shape (B*S, D) or (B*S, t, D)
+        """
+        if self.factorize_space_time:
+            x = self.spatial_attn_agg(x, x_mask)  # (B*S, t, D)
+            x = self.temp_attn_agg(x)  # (B*S, D) or (BS, t, D) if `self.temp_attn_agg` is `Identity`
+        return x
+
+
+
+class TextSynchformer(nn.Module):
+
+    def __init__(self, video: bool = True, audio: bool = False,
+                 text_dim: int = 1024, max_text_seq_len: int = 512, xattn_depth: int = 1):
+        super().__init__()
+        
+        self.video = video
+        self.audio = audio
+        self.text_dim = text_dim
+        self.max_text_seq_len = max_text_seq_len
+        
+        if not video and not audio:
+            raise ValueError('At least one of video or audio should be True.')
+
+        # Use ExtendedMotionFormer directly instead of inheriting from Synchformer
+        if self.video:
+            self.vfeat_extractor = ExtendedMotionFormer(
+                extract_features=True,
+                factorize_space_time=True,
+                agg_space_module='TransformerEncoderLayer',
+                agg_time_module='torch.nn.Identity',
+                add_global_repr=False
+            )
+            
+        if self.audio:
+            self.afeat_extractor = AST(
+                extract_features=True,
+                max_spec_t=66,
+                factorize_freq_time=True,
+                agg_freq_module='TransformerEncoderLayer',
+                agg_time_module='torch.nn.Identity',
+                add_global_repr=False
+            )
+        
+        # Get embedding dimensions from the video feature extractor
+        if self.video:
+            self.embed_dim = self.vfeat_extractor.embed_dim
+            self.num_heads = self.vfeat_extractor.num_heads
+            self.mlp_ratio = self.vfeat_extractor.mlp_ratio
+        else:
+            # Default values if no video
+            self.embed_dim = 768
+            self.num_heads = 12
+            self.mlp_ratio = 4
+        
+        self.text_proj = nn.Sequential(
+            nn.Linear(self.text_dim, self.embed_dim),
+            nn.SiLU(),
+            MLP(self.embed_dim, self.embed_dim * 4)
+        )
+        self.synch_text_cross_blocks = nn.ModuleList([
+            MMCrossAttentionBlock(self.embed_dim, self.num_heads, 
+                                  mlp_ratio=self.mlp_ratio, 
+                                  kernel_size=1, padding=0,
+                                  residual=True)
+            for _ in range(xattn_depth)
+        ])
+    
+    def initialize_weights(self):
+        
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+                    
+        self.text_proj.apply(_basic_init)
+        self.synch_text_cross_blocks.apply(_basic_init)
+        for block in self.synch_text_cross_blocks:
+            nn.init.constant_(block.norm1.weight, 0.0)
+            nn.init.constant_(block.norm1.bias, 0.0)
+            nn.init.constant_(block.ffn.w2.weight, 0.0)
+            
+    
+    def forward(self, data, text_features):
+        video, audio = None, None
+        
+        if self.video and self.audio:
+            video, audio = data
+        elif self.video:
+            video = data
+        elif self.audio:
+            audio = data
+            
+        if self.video and video is not None:
+            video = self.forward_vfeat(video, text_features)
+        if self.audio and audio is not None:
+            audio = self.forward_afeat(audio, text_features)
+        
+        if self.video and self.audio:
+            return video, audio
+        elif self.video:
+            return video
+        else:
+            return audio
+    
+    def forward_vfeat(self, vis, text_f, text_mask):
+        B, S, Tv, C, H, W = vis.shape
+        vis = vis.permute(0, 1, 3, 2, 4, 5)  # (B, S, C, Tv, H, W)
+        
+        # Flatten for processing
+        orig_shape = (B, S, C, Tv, H, W)
+        vis = einops.rearrange(vis, 'B S C Tv H W -> (B S) C Tv H W') # vis.view(B * S, C, Tv, H, W)
+        vis, vis_mask = self.vfeat_extractor.forward_segments_without_aggregation(
+            vis, orig_shape # B*S D t h w , BS t h w
+        )
+        
+        text_f = self.text_proj(text_f)  # (B, text_dim) -> (B, embed_dim)
+        
+        BS, D, t, h, w = vis.shape
+        vis = einops.rearrange(vis, '(B S) D t h w -> B (S t h w) D', B=B, S=S)
+        vis_mask = einops.rearrange(vis_mask, '(B S) t h w -> B (S t h w)', B=B, S=S) \
+            if vis_mask is not None else None
+        for block in self.synch_text_cross_blocks:
+            vis = block(vis, text_f, rot=None, x_mask=vis_mask, context_mask=text_mask)
+        
+        vis = einops.rearrange(vis, 'B (S t h w) D -> (B S) D t h w', B=B, S=S, D=D, t=t, h=h, w=w)
+        vis_mask = einops.rearrange(vis_mask, 'B (S t h w) -> (B S) t h w', B=B, S=S, t=t, h=h, w=w) \
+            if vis_mask is not None else None
+        vis = self.vfeat_extractor.spatiotemporal_aggregation(
+            vis, vis_mask
+        )
+        vis = vis.view(B, S, *vis.shape[1:])
+        
+        return vis
+    
+    def forward_afeat(self, aud):
+        """Forward audio features."""
+        raise NotImplementedError("Audio feature extraction is not implemented in TextSynchformer.")
+        # B, S, F, Ta = aud.shape
+        # aud = aud.permute(0, 1, 3, 2)  # (B, S, Ta, F)
+        # aud, _ = self.afeat_extractor(aud)
+        # return aud
+    
+    def load_state_dict(self, sd: Mapping[str, Any], strict: bool = True):
+        target_keys = (['vfeat_extractor'] if self.video else []) \
+                    + (['afeat_extractor'] if self.audio else []) \
+                    + ['text_proj', 'synch_text_cross_blocks']
+        # discard all entries except vfeat_extractor / afeat_extractor
+        sd = {k: v for k, v in sd.items() if any(k.startswith(tk) 
+                                                 for tk in target_keys)}
+        
+
+        return super().load_state_dict(sd, strict)
+    

selva_core/model/transformer_layers.py

@@ -0,0 +1,426 @@
+from typing import Optional
+from inspect import isfunction
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+from selva_core.ext.rotary_embeddings import apply_rope
+from selva_core.model.low_level import MLP, ChannelLastConv1d, ConvMLP
+
+
+def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor):
+    return x * (1 + scale) + shift
+
+
+def attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+              attn_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
+    # training will crash without these contiguous calls and the CUDNN limitation
+    # I believe this is related to https://github.com/pytorch/pytorch/issues/133974
+    # unresolved at the time of writing
+    q = q.contiguous()
+    k = k.contiguous()
+    v = v.contiguous()
+    if attn_mask is not None:
+        attn_mask = attn_mask.contiguous()
+    out = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
+    # out = rearrange(out, 'b h n d -> b n (h d)').contiguous()
+    b, h, n, d_head = out.shape
+    out = out.permute(0, 2, 1, 3)  # Shape: (b, n, h, d_head)
+    # Using reshape, which can handle non-contiguous tensors by copying if necessary
+    out = out.reshape(b, n, h * d_head) # Shape: (b, n, h * d_head)
+    # Ensure the final output is contiguous, similar to the original code's intent
+    out = out.contiguous()
+    return out
+
+def attention_debug(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+              attn_mask: Optional[torch.Tensor] = None,
+              layer_idx: int = -1) -> None:
+    # training will crash without these contiguous calls and the CUDNN limitation
+    # I believe this is related to https://github.com/pytorch/pytorch/issues/133974
+    # unresolved at the time of writing
+    q = q.contiguous()
+    k = k.contiguous()
+    v = v.contiguous()
+    if attn_mask is not None:
+        attn_mask = attn_mask.contiguous()
+    # out = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
+    
+    # debug attn map
+    import math
+    scale_factor = 1 / math.sqrt(q.size(-1))
+    L, S = q.size(-2), k.size(-2)
+    attn_bias = torch.zeros(q.shape[0], q.shape[1], L, S, dtype=q.dtype, device=q.device)
+    if attn_mask is not None:
+        if attn_mask.dtype == torch.bool:
+            attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
+        else:
+            attn_bias = attn_mask + attn_bias
+    attn_weight = q @ k.transpose(-2, -1) * scale_factor
+    attn_weight += attn_bias
+    torch.save(attn_weight.clone().cpu(), f'./debug_attn_weight_layer{layer_idx}_unnorm.pt')
+    # normalize
+    attn_weight = torch.softmax(attn_weight, dim=-1)
+    torch.save(attn_weight.clone().cpu(), f'./debug_attn_weight_layer{layer_idx}.pt')
+
+
+def create_mask(q_shape, k_shape, device, q_mask=None, k_mask=None):
+    def default(val, d):
+        return val if val is not None else (d() if isfunction(d) else d)
+    b, i, j, device = q_shape[0], q_shape[-2], k_shape[-2], device
+    q_mask = default(q_mask, torch.ones((b, i), device=device, dtype=torch.bool))
+    k_mask = default(k_mask, torch.ones((b, j), device=device, dtype=torch.bool))
+    attn_mask = rearrange(q_mask, 'b i -> b 1 i 1') * rearrange(k_mask, 'b j -> b 1 1 j')
+    return attn_mask
+
+
+class SelfAttention(nn.Module):
+
+    def __init__(self, dim: int, nheads: int):
+        super().__init__()
+        self.dim = dim
+        self.nheads = nheads
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=True)
+        self.q_norm = nn.RMSNorm(dim // nheads)
+        self.k_norm = nn.RMSNorm(dim // nheads)
+
+        self.split_into_heads = Rearrange('b n (h d j) -> b h n d j',
+                                          h=nheads,
+                                          d=dim // nheads,
+                                          j=3)
+
+    def pre_attention(
+            self, x: torch.Tensor,
+            rot: Optional[torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        # x: batch_size * n_tokens * n_channels
+        qkv = self.qkv(x)
+        q, k, v = self.split_into_heads(qkv).chunk(3, dim=-1)
+        q = q.squeeze(-1)
+        k = k.squeeze(-1)
+        v = v.squeeze(-1)
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        if rot is not None:
+            q = apply_rope(q, rot)
+            k = apply_rope(k, rot)
+
+        return q, k, v
+
+    def forward(
+            self,
+            x: torch.Tensor,  # batch_size * n_tokens * n_channels
+            q_mask: Optional[torch.Tensor] = None,
+            k_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        q, v, k = self.pre_attention(x)
+        if q_mask is not None or k_mask is not None:
+            attn_mask = create_mask(q.shape, k.shape, q.device, 
+                                    q_mask=q_mask, k_mask=k_mask)
+        else:
+            attn_mask = None
+        out = attention(q, k, v, attn_mask)
+        return out
+
+
+class CrossAttention(nn.Module):
+
+    def __init__(self, dim: int, nheads: int):
+        """
+        Args:
+            dim (int): Input dimension.
+            nheads (int): Number of attention heads.
+            
+        Attributes:
+            q_proj (Linear): Linear transformation for the query.
+            kv_proj (Linear): Linear transformation for the key and value.
+            q_norm (RMSNorm): Layer normalization for the query.
+            k_norm (RMSNorm): Layer normalization for the key.
+            split_into_heads (Rearrange): Rearrange layer to split the input into heads.
+        """
+        super().__init__()
+        self.dim = dim
+        self.nheads = nheads
+
+        self.q_proj = nn.Linear(dim, dim, bias=True)
+        self.kv_proj = nn.Linear(dim, dim * 2, bias=True)
+        self.q_norm = nn.RMSNorm(dim // nheads)
+        self.k_norm = nn.RMSNorm(dim // nheads)
+
+        self.split_q_into_heads = Rearrange('b n (h d) -> b h n d',
+                          h=nheads,
+                          d=dim // nheads)
+        self.split_kv_into_heads = Rearrange('b n (h d j) -> b h n d j',
+                          h=nheads,
+                          d=dim // nheads,
+                          j=2)
+
+    def pre_attention(
+            self, x: torch.Tensor, c: torch.Tensor,
+            rot: Optional[torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        # x: batch_size * n_tokens * n_channels
+        # c: batch_size * n_cond_tokens * n_channels
+        q = self.q_proj(x)
+        kv = self.kv_proj(c)
+        q = self.split_q_into_heads(q)
+        k, v = self.split_kv_into_heads(kv).chunk(2, dim=-1)
+        k = k.squeeze(-1)
+        v = v.squeeze(-1)
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+        
+        if rot is not None:
+            q = apply_rope(q, rot)
+
+        return q, k, v
+
+    def forward(
+            self,
+            x: torch.Tensor,  # batch_size * n_tokens * n_channels
+            c: torch.Tensor,  # batch_size * n_cond_tokens * n_channels
+            context_mask: Optional[torch.Tensor] = None,
+            rot: Optional[torch.Tensor] = None
+        ) -> torch.Tensor:
+        q, k, v = self.pre_attention(x, c, rot)
+        if context_mask is not None:
+            attn_mask = create_mask(q.shape, k.shape, q.device, k_mask=context_mask)
+        else:
+            attn_mask = None
+        out = attention(q, k, v, attn_mask)
+        return out
+
+
+class MMCrossAttentionBlock(nn.Module):
+
+    def __init__(self,
+                 dim: int,
+                 nhead: int,
+                 mlp_ratio: float = 4.0,
+                #  pre_only: bool = False,
+                 kernel_size: int = 7,
+                 padding: int = 3,
+                 residual: bool = True):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=True)
+        self.attn = CrossAttention(dim, nhead)
+
+        if kernel_size == 1:
+            self.linear1 = nn.Linear(dim, dim)
+        else:
+            self.linear1 = ChannelLastConv1d(dim, dim, kernel_size=kernel_size, padding=padding)
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=True)
+
+        if kernel_size == 1:
+            self.ffn = MLP(dim, int(dim * mlp_ratio))
+        else:
+            self.ffn = ConvMLP(dim,
+                                int(dim * mlp_ratio),
+                                kernel_size=kernel_size,
+                                padding=padding)
+
+        self.residual = residual
+
+    def pre_attention(self, x: torch.Tensor, c: torch.Tensor, rot: Optional[torch.Tensor]):
+        # x: BS * N * D
+        # cond: BS * D
+        
+        # if self.pre_only:
+        #     (shift_msa, scale_msa) = modulation.chunk(2, dim=-1)
+        #     gate_msa = shift_mlp = scale_mlp = gate_mlp = None
+        # else:
+        #     (shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp,
+        #      gate_mlp) = modulation.chunk(6, dim=-1)
+
+        # x = self.norm1(x)
+        q, k, v = self.attn.pre_attention(x, c, rot)
+        return (q, k, v)
+
+    def post_attention(self, x: torch.Tensor, attn_out: torch.Tensor):
+        # if self.pre_only:
+        #     return x
+
+        # (gate_msa, shift_mlp, scale_mlp, gate_mlp) = c
+        if self.residual:
+            x = x + self.norm1(self.linear1(attn_out)) # * gate_msa
+            # https://github.com/haidog-yaqub/EzAudio/blob/2eb0bd90013584c6e28a6c14ec28b935f1e78de5/src/models/blocks.py#L158
+            # https://github.com/huggingface/diffusers/blob/07dd6f8c0e267662f62c39cd8334c2b5d157ab39/src/diffusers/models/transformers/transformer_flux.py#L170
+            # https://github.com/Stability-AI/stablediffusion/blob/cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf/ldm/modules/attention.py#L274
+        else:
+            x = self.norm1(self.linear1(attn_out))
+        r = self.norm2(x)
+        x = x + self.ffn(r)
+
+        return x
+
+    def forward(self, x: torch.Tensor, cond: torch.Tensor,
+                rot: Optional[torch.Tensor], 
+                x_mask: Optional[torch.Tensor] = None,
+                context_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
+        # x: BS * N * D
+        # cond: BS * D
+        q, k, v = self.pre_attention(x, cond, rot)
+        if x_mask is not None or context_mask is not None:
+            attn_mask = create_mask(q.shape, k.shape, q.device, q_mask=x_mask, k_mask=context_mask)
+        else:
+            attn_mask = None
+        attn_out = attention(q, k, v, attn_mask=attn_mask)
+        x = self.post_attention(x, attn_out)
+
+        return x
+
+
+class MMDitSingleBlock(nn.Module):
+
+    def __init__(self,
+                 dim: int,
+                 nhead: int,
+                 mlp_ratio: float = 4.0,
+                 pre_only: bool = False,
+                 kernel_size: int = 7,
+                 padding: int = 3):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False)
+        self.attn = SelfAttention(dim, nhead)
+
+        self.pre_only = pre_only
+        if pre_only:
+            self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 2 * dim, bias=True))
+        else:
+            if kernel_size == 1:
+                self.linear1 = nn.Linear(dim, dim)
+            else:
+                self.linear1 = ChannelLastConv1d(dim, dim, kernel_size=kernel_size, padding=padding)
+            self.norm2 = nn.LayerNorm(dim, elementwise_affine=False)
+
+            if kernel_size == 1:
+                self.ffn = MLP(dim, int(dim * mlp_ratio))
+            else:
+                self.ffn = ConvMLP(dim,
+                                   int(dim * mlp_ratio),
+                                   kernel_size=kernel_size,
+                                   padding=padding)
+
+            self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 6 * dim, bias=True))
+
+    def pre_attention(self, x: torch.Tensor, c: torch.Tensor, rot: Optional[torch.Tensor]):
+        # x: BS * N * D
+        # cond: BS * D
+        modulation = self.adaLN_modulation(c)
+        if self.pre_only:
+            (shift_msa, scale_msa) = modulation.chunk(2, dim=-1)
+            gate_msa = shift_mlp = scale_mlp = gate_mlp = None
+        else:
+            (shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp,
+             gate_mlp) = modulation.chunk(6, dim=-1)
+
+        x = modulate(self.norm1(x), shift_msa, scale_msa)
+        q, k, v = self.attn.pre_attention(x, rot)
+        return (q, k, v), (gate_msa, shift_mlp, scale_mlp, gate_mlp)
+
+    def post_attention(self, x: torch.Tensor, attn_out: torch.Tensor, c: tuple[torch.Tensor]):
+        if self.pre_only:
+            return x
+
+        (gate_msa, shift_mlp, scale_mlp, gate_mlp) = c
+        x = x + self.linear1(attn_out) * gate_msa
+        r = modulate(self.norm2(x), shift_mlp, scale_mlp)
+        x = x + self.ffn(r) * gate_mlp
+
+        return x
+
+    def forward(self, x: torch.Tensor, cond: torch.Tensor,
+                rot: Optional[torch.Tensor]) -> torch.Tensor:
+        # x: BS * N * D
+        # cond: BS * D
+        x_qkv, x_conditions = self.pre_attention(x, cond, rot)
+        attn_out = attention(*x_qkv)
+        x = self.post_attention(x, attn_out, x_conditions)
+
+        return x
+
+
+class JointBlock(nn.Module):
+
+    def __init__(self, dim: int, nhead: int, mlp_ratio: float = 4.0, pre_only: bool = False):
+        super().__init__()
+        self.pre_only = pre_only
+        self.latent_block = MMDitSingleBlock(dim,
+                                             nhead,
+                                             mlp_ratio,
+                                             pre_only=False,
+                                             kernel_size=3,
+                                             padding=1)
+        self.clip_block = MMDitSingleBlock(dim,
+                                           nhead,
+                                           mlp_ratio,
+                                           pre_only=pre_only,
+                                           kernel_size=3,
+                                           padding=1)
+        self.text_block = MMDitSingleBlock(dim, nhead, mlp_ratio, pre_only=pre_only, kernel_size=1)
+
+    def forward(self, latent: torch.Tensor, clip_f: torch.Tensor, text_f: torch.Tensor,
+                global_c: torch.Tensor, extended_c: torch.Tensor, latent_rot: torch.Tensor,
+                clip_rot: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        # latent: BS * N1 * D
+        # clip_f: BS * N2 * D
+        # c: BS * (1/N) * D
+        x_qkv, x_mod = self.latent_block.pre_attention(latent, extended_c, latent_rot)
+        c_qkv, c_mod = self.clip_block.pre_attention(clip_f, global_c, clip_rot)
+        t_qkv, t_mod = self.text_block.pre_attention(text_f, global_c, rot=None)
+
+        latent_len = latent.shape[1]
+        clip_len = clip_f.shape[1]
+        text_len = text_f.shape[1]
+
+        joint_qkv = [torch.cat([x_qkv[i], c_qkv[i], t_qkv[i]], dim=2) for i in range(3)]
+
+        attn_out = attention(*joint_qkv)
+        x_attn_out = attn_out[:, :latent_len]
+        c_attn_out = attn_out[:, latent_len:latent_len + clip_len]
+        t_attn_out = attn_out[:, latent_len + clip_len:]
+
+        latent = self.latent_block.post_attention(latent, x_attn_out, x_mod)
+        if not self.pre_only:
+            clip_f = self.clip_block.post_attention(clip_f, c_attn_out, c_mod)
+            text_f = self.text_block.post_attention(text_f, t_attn_out, t_mod)
+
+        return latent, clip_f, text_f
+    
+    def forward_debug(self, latent: torch.Tensor, clip_f: torch.Tensor, text_f: torch.Tensor,
+                global_c: torch.Tensor, extended_c: torch.Tensor, latent_rot: torch.Tensor,
+                clip_rot: torch.Tensor,
+                layer_idx: int = -1,
+                ) -> None:
+        # latent: BS * N1 * D
+        # clip_f: BS * N2 * D
+        # c: BS * (1/N) * D
+        x_qkv, x_mod = self.latent_block.pre_attention(latent, extended_c, latent_rot)
+        c_qkv, c_mod = self.clip_block.pre_attention(clip_f, global_c, clip_rot)
+        t_qkv, t_mod = self.text_block.pre_attention(text_f, global_c, rot=None)
+
+        latent_len = latent.shape[1]
+        clip_len = clip_f.shape[1]
+        text_len = text_f.shape[1]
+
+        joint_qkv = [torch.cat([x_qkv[i], c_qkv[i], t_qkv[i]], dim=2) for i in range(3)]
+
+        attn_out = attention_debug(*joint_qkv, layer_idx=layer_idx)
+        return None
+
+
+class FinalBlock(nn.Module):
+
+    def __init__(self, dim, out_dim):
+        super().__init__()
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(dim, 2 * dim, bias=True))
+        self.norm = nn.LayerNorm(dim, elementwise_affine=False)
+        self.conv = ChannelLastConv1d(dim, out_dim, kernel_size=7, padding=3)
+
+    def forward(self, latent, c):
+        shift, scale = self.adaLN_modulation(c).chunk(2, dim=-1)
+        latent = modulate(self.norm(latent), shift, scale)
+        latent = self.conv(latent)
+        return latent

selva_core/model/utils/distributions.py

@@ -0,0 +1,46 @@
+from typing import Optional
+
+import numpy as np
+import torch
+
+
+class DiagonalGaussianDistribution:
+
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self, rng: Optional[torch.Generator] = None):
+        # x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+
+        r = torch.empty_like(self.mean).normal_(generator=rng)
+        x = self.mean + self.std * r
+
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+
+                return 0.5 * torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar
+            else:
+                return 0.5 * (torch.pow(self.mean - other.mean, 2) / other.var +
+                              self.var / other.var - 1.0 - self.logvar + other.logvar)
+
+    def nll(self, sample, dims=[1, 2, 3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+                               dim=dims)
+
+    def mode(self):
+        return self.mean

selva_core/model/utils/factory.py

@@ -0,0 +1,18 @@
+import torch
+
+from selva_core.utils.misc import instantiate_from_config
+from selva_core.model.networks_video_enc import TextSynch as TextSynchVideoEnc
+from selva_core.model.networks_generator import MMAudio
+
+
+_MODEL_ZOO = (TextSynchVideoEnc, MMAudio)
+
+
+def create_model_from_factory(factory_path: str, name: str, **kwargs) -> torch.nn.Module:
+    """
+    Dynamically imports and calls a model factory function.
+    """
+    params = {'name': name, **kwargs}
+    model = instantiate_from_config(factory_path, params)
+    assert isinstance(model, _MODEL_ZOO), f"Model {type(model)} is not a valid model type."
+    return model

selva_core/model/utils/features_utils.py

@@ -0,0 +1,192 @@
+from typing import Literal, Optional
+
+import open_clip
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from open_clip import create_model_from_pretrained
+from torchvision.transforms import Normalize
+from transformers import T5TokenizerFast, T5EncoderModel
+
+from selva_core.ext.autoencoder import AutoEncoderModule
+from selva_core.ext.mel_converter import get_mel_converter
+from selva_core.ext.synchformer import Synchformer
+from selva_core.model.utils.distributions import DiagonalGaussianDistribution
+from selva_core.utils.transforms import generate_multiple_segments
+
+
+def patch_clip(clip_model):
+    # a hack to make it output last hidden states
+    # https://github.com/mlfoundations/open_clip/blob/fc5a37b72d705f760ebbc7915b84729816ed471f/src/open_clip/model.py#L269
+    def new_encode_text(self, text, normalize: bool = False):
+        cast_dtype = self.transformer.get_cast_dtype()
+
+        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]
+
+        x = x + self.positional_embedding.to(cast_dtype)
+        x = self.transformer(x, attn_mask=self.attn_mask)
+        x = self.ln_final(x)  # [batch_size, n_ctx, transformer.width]
+        return F.normalize(x, dim=-1) if normalize else x
+
+    clip_model.encode_text = new_encode_text.__get__(clip_model)
+    return clip_model
+
+
+class FeaturesUtils(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        tod_vae_ckpt: Optional[str] = None,
+        bigvgan_vocoder_ckpt: Optional[str] = None,
+        synchformer_ckpt: Optional[str] = None,
+        enable_conditions: bool = True,
+        mode=Literal['16k', '44k'],
+        need_vae_encoder: bool = True,
+    ):
+        super().__init__()
+
+        if enable_conditions:
+            self.clip_model = create_model_from_pretrained('hf-hub:apple/DFN5B-CLIP-ViT-H-14-384',
+                                                           return_transform=False)
+            self.clip_preprocess = Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
+                                             std=[0.26862954, 0.26130258, 0.27577711])
+            self.clip_model = patch_clip(self.clip_model)
+            self.tokenizer_clip = open_clip.get_tokenizer('ViT-H-14-378-quickgelu')  # same as 'ViT-H-14'
+
+            self.synchformer = Synchformer(video=True, audio=False)
+            self.synchformer.load_state_dict(
+                torch.load(synchformer_ckpt, weights_only=True, map_location='cpu'))
+            
+            self.text_encoder_t5 = T5EncoderModel.from_pretrained('google/flan-t5-base')
+            self.tokenizer_t5 = T5TokenizerFast.from_pretrained('google/flan-t5-base')
+        else:
+            self.clip_model = None
+            self.synchformer = None
+            self.tokenizer_clip = None
+            self.text_encoder_t5 = None
+            self.tokenizer_t5 = None
+            
+        if tod_vae_ckpt is not None:
+            self.mel_converter = get_mel_converter(mode)
+            self.tod = AutoEncoderModule(vae_ckpt_path=tod_vae_ckpt,
+                                         vocoder_ckpt_path=bigvgan_vocoder_ckpt,
+                                         mode=mode,
+                                         need_vae_encoder=need_vae_encoder)
+        else:
+            self.tod = None
+
+    def compile(self):
+        if self.clip_model is not None:
+            self.clip_model.encode_image = torch.compile(self.clip_model.encode_image)
+            self.clip_model.encode_text = torch.compile(self.clip_model.encode_text)
+        if self.synchformer is not None:
+            self.synchformer = torch.compile(self.synchformer)
+            self.synchformer.forward_vfeat = torch.compile(self.synchformer.forward_vfeat)
+        if self.text_encoder_t5 is not None:
+            self.text_encoder_t5.forward = torch.compile(self.text_encoder_t5.forward)
+        self.decode = torch.compile(self.decode)
+        self.vocode = torch.compile(self.vocode)
+
+    def train(self, mode: bool) -> None:
+        return super().train(False)
+
+    @torch.inference_mode()
+    def encode_video_with_clip(self, x: torch.Tensor, batch_size: int = -1) -> torch.Tensor:
+        assert self.clip_model is not None, 'CLIP is not loaded'
+        # x: (B, T, C, H, W) H/W: 384
+        b, t, c, h, w = x.shape
+        assert c == 3 and h == 384 and w == 384
+        x = self.clip_preprocess(x)
+        x = rearrange(x, 'b t c h w -> (b t) c h w')
+        outputs = []
+        if batch_size < 0:
+            batch_size = b * t
+        for i in range(0, b * t, batch_size):
+            outputs.append(self.clip_model.encode_image(x[i:i + batch_size], normalize=True))
+        x = torch.cat(outputs, dim=0)
+        # x = self.clip_model.encode_image(x, normalize=True)
+        x = rearrange(x, '(b t) d -> b t d', b=b)
+        return x
+
+    @torch.inference_mode()
+    def encode_video_with_sync(self, x: torch.Tensor, batch_size: int = -1) -> torch.Tensor:
+        assert self.synchformer is not None, 'Synchformer is not loaded'
+        # x: (B, T, C, H, W) H/W: 384
+
+        b, t, c, h, w = x.shape
+        assert c == 3 and h == 224 and w == 224
+
+        # partition the video
+        segment_size = 16
+        step_size = 8
+        x = generate_multiple_segments(x, segment_size, step_size) # (B, S, T, C, H, W)
+        num_segments = x.shape[1]
+
+        outputs = []
+        if batch_size < 0:
+            batch_size = b
+        x = rearrange(x, 'b s t c h w -> (b s) 1 t c h w')
+        for i in range(0, b * num_segments, batch_size):
+            outputs.append(self.synchformer.forward_vfeat(x[i:i + batch_size]))
+        x = torch.cat(outputs, dim=0)
+        x = rearrange(x, '(b s) 1 t d -> b (s t) d', b=b)
+        return x
+
+    @torch.inference_mode()
+    def encode_text_clip(self, text: list[str]) -> torch.Tensor:
+        assert self.clip_model is not None, 'CLIP is not loaded'
+        assert self.tokenizer_clip is not None, 'Tokenizer is not loaded'
+        # x: (B, L)
+        tokens = self.tokenizer_clip(text).to(self.device)
+        return self.clip_model.encode_text(tokens, normalize=True)
+    
+    @torch.inference_mode()
+    def encode_text_t5(self, text: list[str]) -> torch.Tensor:
+        device = self.text_encoder_t5.device
+        batch = self.tokenizer_t5(
+            text,
+            max_length=self.tokenizer_t5.model_max_length,
+            padding=True,
+            truncation=True,
+            return_tensors="pt",
+        )
+        input_ids, attention_mask = batch.input_ids.to(device), batch.attention_mask.to(
+            device
+        )
+
+        encoder_hidden_states = self.text_encoder_t5(
+            input_ids=input_ids, attention_mask=attention_mask
+        ).last_hidden_state # (B, L, D)
+
+        boolean_encoder_mask = (attention_mask == 1).to(device) # (B, L)
+        
+        return encoder_hidden_states, boolean_encoder_mask
+
+    @torch.inference_mode()
+    def encode_audio(self, x) -> DiagonalGaussianDistribution:
+        assert self.tod is not None, 'VAE is not loaded'
+        # x: (B * L)
+        mel = self.mel_converter(x)
+        dist = self.tod.encode(mel)
+
+        return dist
+    
+    @torch.inference_mode()
+    def vocode(self, mel: torch.Tensor) -> torch.Tensor:
+        assert self.tod is not None, 'VAE is not loaded'
+        return self.tod.vocode(mel)
+
+    @torch.inference_mode()
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        assert self.tod is not None, 'VAE is not loaded'
+        return self.tod.decode(z.transpose(1, 2))
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    @property
+    def dtype(self):
+        return next(self.parameters()).dtype

selva_core/model/utils/parameter_groups.py

@@ -0,0 +1,39 @@
+import logging
+
+log = logging.getLogger()
+
+
+def get_parameter_groups(model, cfg, print_log=False):
+    """
+    Assign different weight decays and learning rates to different parameters.
+    Returns a parameter group which can be passed to the optimizer.
+    """
+    weight_decay = cfg.weight_decay
+    base_lr = cfg.learning_rate
+
+    params = []
+
+    # inspired by detectron2
+    memo = set()
+    for name, param in model.named_parameters():
+        if not param.requires_grad:
+            continue
+        # Avoid duplicating parameters
+        if param in memo:
+            continue
+        memo.add(param)
+
+        if name.startswith('module'):
+            name = name[7:]
+
+        params.append(param)
+
+    parameter_groups = [
+        {
+            'params': params,
+            'lr': base_lr,
+            'weight_decay': weight_decay
+        },
+    ]
+
+    return parameter_groups

selva_core/model/utils/sample_utils.py

@@ -0,0 +1,12 @@
+from typing import Optional
+
+import torch
+
+
+def log_normal_sample(x: torch.Tensor,
+                      generator: Optional[torch.Generator] = None,
+                      m: float = 0.0,
+                      s: float = 1.0) -> torch.Tensor:
+    bs = x.shape[0]
+    s = torch.randn(bs, device=x.device, generator=generator) * s + m
+    return torch.sigmoid(s)

selva_core/utils/dist_utils.py

@@ -0,0 +1,17 @@
+import os
+from logging import Logger
+
+from selva_core.utils.logger import TensorboardLogger
+
+local_rank = int(os.environ['LOCAL_RANK']) if 'LOCAL_RANK' in os.environ else 0
+world_size = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1
+
+
+def info_if_rank_zero(logger: Logger, msg: str):
+    if local_rank == 0:
+        logger.info(msg)
+
+
+def string_if_rank_zero(logger: TensorboardLogger, tag: str, msg: str):
+    if local_rank == 0:
+        logger.log_string(tag, msg)

selva_core/utils/download_utils.py

@@ -0,0 +1,109 @@
+import hashlib
+import logging
+from pathlib import Path
+
+import requests
+from tqdm import tqdm
+
+log = logging.getLogger()
+
+links = [
+    {
+        'name': 'video_enc_sup_5.pth',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/weights/video_enc_sup_5.pth',
+        'md5': 'ff09a6dc36148536ee4db97eba081d05'
+    },
+    {
+        'name': 'generator_small_16k_sup_5.pth',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/weights/generator_small_16k_sup_5.pth',
+        'md5': '1cb0f0deec52de37f67b1fd9965337d0'
+    },
+    {
+        'name': 'generator_small_44k_sup_5.pth',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/weights/generator_small_44k_sup_5.pth',
+        'md5': 'd4df8569624093ac80af99b8b7434525'
+    },
+    {
+        'name': 'generator_medium_44k_sup_5.pth',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/weights/generator_medium_44k_sup_5.pth',
+        'md5': 'e9157e62b4863ad306e89e8f3a587748'
+    },
+    {
+        'name': 'generator_large_44k_sup_5.pth',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/weights/generator_large_44k_sup_5.pth',
+        'md5': 'ab3db08b124d3aaa53eb7a1f52f1fb3f'
+    },
+    {
+        'name': 'v1-16.pth',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/ext_weights/v1-16.pth',
+        'md5': '69f56803f59a549a1a507c93859fd4d7'
+    },
+    {
+        'name': 'best_netG.pt',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/ext_weights/best_netG.pt',
+        'md5': 'eeaf372a38a9c31c362120aba2dde292'
+    },
+    {
+        'name': 'v1-44.pth',
+        'url': 'https://huggingface.co/jnwnlee/SelVA/resolve/main/ext_weights/v1-44.pth',
+        'md5': 'fab020275fa44c6589820ce025191600'
+    },
+    {
+        'name': 'synchformer_state_dict.pth',
+        'url':
+        'https://huggingface.co/jnwnlee/SelVA/resolve/main/ext_weights/synchformer_state_dict.pth',
+        'md5': '5b2f5594b0730f70e41e549b7c94390c'
+    },
+    {
+        'name': 'mmaudio_small_16k.pth',
+        'url': 'https://huggingface.co/hkchengrex/MMAudio/resolve/main/weights/mmaudio_small_16k.pth',
+        'md5': 'af93cde404179f58e3919ac085b8033b',
+    },
+    {
+        'name': 'mmaudio_small_44k.pth',
+        'url': 'https://huggingface.co/hkchengrex/MMAudio/resolve/main/weights/mmaudio_small_44k.pth',
+        'md5': 'babd74c884783d13701ea2820a5f5b6d',
+    },
+    {
+        'name': 'mmaudio_medium_44k.pth',
+        'url': 'https://huggingface.co/hkchengrex/MMAudio/resolve/main/weights/mmaudio_medium_44k.pth',
+        'md5': '5a56b6665e45a1e65ada534defa903d0',
+    },
+    {
+        'name': 'mmaudio_large_44k.pth',
+        'url': 'https://huggingface.co/hkchengrex/MMAudio/resolve/main/weights/mmaudio_large_44k.pth',
+        'md5': 'fed96c325a6785b85ce75ae1aafd2673'
+    },
+    {
+        'name': 'mmaudio_large_44k_v2.pth',
+        'url': 'https://huggingface.co/hkchengrex/MMAudio/resolve/main/weights/mmaudio_large_44k_v2.pth',
+        'md5': '01ad4464f049b2d7efdaa4c1a59b8dfe'
+    },
+]
+
+
+def download_model_if_needed(model_path: Path):
+    base_name = model_path.name
+
+    for link in links:
+        if link['name'] == base_name:
+            target_link = link
+            break
+    else:
+        raise ValueError(f'No link found for {base_name}')
+
+    model_path.parent.mkdir(parents=True, exist_ok=True)
+    if not model_path.exists() or hashlib.md5(open(model_path,
+                                                   'rb').read()).hexdigest() != target_link['md5']:
+        log.info(f'Downloading {base_name} to {model_path}...')
+        r = requests.get(target_link['url'], stream=True)
+        total_size = int(r.headers.get('content-length', 0))
+        block_size = 1024
+        t = tqdm(total=total_size, unit='iB', unit_scale=True)
+        with open(model_path, 'wb') as f:
+            for data in r.iter_content(block_size):
+                t.update(len(data))
+                f.write(data)
+        t.close()
+        if total_size != 0 and t.n != total_size:
+            raise RuntimeError('Error while downloading %s' % base_name)

selva_core/utils/email_utils.py

@@ -0,0 +1,50 @@
+import logging
+import os
+from datetime import datetime
+
+import requests
+from dotenv import load_dotenv
+from pytz import timezone
+
+from selva_core.utils.timezone import my_timezone
+
+_source = 'USE YOURS'
+_target = 'USE YOURS'
+
+log = logging.getLogger()
+
+_fmt = "%Y-%m-%d %H:%M:%S %Z%z"
+
+
+class EmailSender:
+
+    def __init__(self, exp_id: str, enable: bool):
+        self.exp_id = exp_id
+        self.enable = enable
+        if enable:
+            load_dotenv()
+            self.MAILGUN_API_KEY = os.getenv('MAILGUN_API_KEY')
+            if self.MAILGUN_API_KEY is None:
+                log.warning('MAILGUN_API_KEY is not set')
+                self.enable = False
+
+    def send(self, subject, content):
+        if self.enable:
+            subject = str(subject)
+            content = str(content)
+            try:
+                return requests.post(f'https://api.mailgun.net/v3/{_source}/messages',
+                                     auth=('api', self.MAILGUN_API_KEY),
+                                     data={
+                                         'from':
+                                         f'<agent name>🤖 <mailgun@{_source}>',
+                                         'to': [f'{_target}'],
+                                         'subject':
+                                         f'[{self.exp_id}] {subject}',
+                                         'text':
+                                         ('\n\n' + content + '\n\n<your sign off>\n' +
+                                          datetime.now(timezone(my_timezone)).strftime(_fmt)),
+                                     },
+                                     timeout=20)
+            except Exception as e:
+                log.error(f'Failed to send email: {e}')

selva_core/utils/eval_utils.py

@@ -0,0 +1,277 @@
+import dataclasses
+import logging
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+from colorlog import ColoredFormatter
+from PIL import Image
+from torchvision.transforms import v2
+
+from selva_core.data.av_utils import ImageInfo, VideoInfo, read_frames, reencode_with_audio
+from selva_core.model.flow_matching import FlowMatching
+from selva_core.model.networks_video_enc import TextSynch
+from selva_core.model.networks_generator import MMAudio
+from selva_core.model.sequence_config import CONFIG_16K, CONFIG_44K, SequenceConfig
+from selva_core.model.utils.features_utils import FeaturesUtils
+from selva_core.utils.download_utils import download_model_if_needed
+
+log = logging.getLogger()
+
+
+@dataclasses.dataclass
+class ModelConfig:
+    model_name: str
+    model_video_enc_path: Path
+    model_generator_path: Path
+    mode: str
+    vae_path: Path
+    bigvgan_16k_path: Optional[Path]
+    synchformer_ckpt: Path = Path('./ext_weights/synchformer_state_dict.pth')
+
+    @property
+    def seq_cfg(self) -> SequenceConfig:
+        if self.mode == '16k':
+            return CONFIG_16K
+        elif self.mode == '44k':
+            return CONFIG_44K
+
+    def download_if_needed(self):
+        download_model_if_needed(self.model_video_enc_path)
+        download_model_if_needed(self.model_generator_path)
+        download_model_if_needed(self.vae_path)
+        if self.bigvgan_16k_path is not None:
+            download_model_if_needed(self.bigvgan_16k_path)
+        download_model_if_needed(self.synchformer_ckpt)
+    
+    def download_video_enc_if_needed(self):
+        download_model_if_needed(self.model_video_enc_path)
+    
+    def download_generator_if_needed(self):
+        download_model_if_needed(self.model_generator_path)
+    
+    def download_external_modules_if_needed(self):
+        download_model_if_needed(self.synchformer_ckpt)
+        download_model_if_needed(self.vae_path)
+        if self.bigvgan_16k_path is not None:
+            download_model_if_needed(self.bigvgan_16k_path)
+
+
+small_16k = ModelConfig(model_name='small_16k',
+                        model_video_enc_path=Path('./weights/video_enc_sup_5.pth'),
+                        model_generator_path=Path('./weights/generator_small_16k_sup_5.pth'),
+                        vae_path=Path('./ext_weights/v1-16.pth'),
+                        bigvgan_16k_path=Path('./ext_weights/best_netG.pt'),
+                        mode='16k')
+small_44k = ModelConfig(model_name='small_44k',
+                        model_video_enc_path=Path('./weights/video_enc_sup_5.pth'),
+                        model_generator_path=Path('./weights/generator_small_44k_sup_5.pth'),
+                        vae_path=Path('./ext_weights/v1-44.pth'),
+                        bigvgan_16k_path=None,
+                        mode='44k')
+medium_44k = ModelConfig(model_name='medium_44k',
+                         model_video_enc_path=Path('./weights/video_enc_sup_5.pth'),
+                         model_generator_path=Path('./weights/generator_medium_44k_sup_5.pth'),
+                         vae_path=Path('./ext_weights/v1-44.pth'),
+                         bigvgan_16k_path=None,
+                         mode='44k')
+large_44k = ModelConfig(model_name='large_44k',
+                        model_video_enc_path=Path('./weights/video_enc_sup_5.pth'),
+                        model_generator_path=Path('./weights/generator_large_44k_sup_5.pth'),
+                        vae_path=Path('./ext_weights/v1-44.pth'),
+                        bigvgan_16k_path=None,
+                        mode='44k')
+all_model_cfg: dict[str, ModelConfig] = {
+    'small_16k': small_16k,
+    'small_44k': small_44k,
+    'medium_44k': medium_44k,
+    'large_44k': large_44k,
+}
+
+
+def generate(
+    clip_video: Optional[torch.Tensor],
+    sync_video: Optional[torch.Tensor],
+    text: Optional[list[str]],
+    *,
+    negative_text: Optional[list[str]] = None,
+    feature_utils: FeaturesUtils,
+    net_video_enc: TextSynch,
+    net_generator: MMAudio,
+    fm: FlowMatching,
+    rng: torch.Generator,
+    cfg_strength: float,
+    clip_batch_size_multiplier: int = 40,
+    sync_batch_size_multiplier: int = 40,
+    image_input: bool = False,
+) -> torch.Tensor:
+    device = feature_utils.device
+    dtype = feature_utils.dtype
+
+    bs = len(text)
+    if text is not None:
+        text_features_clip = feature_utils.encode_text_clip(text)
+        text_features_flant5, text_mask_flant5 = feature_utils.encode_text_t5(text)
+    else:
+        text_features_clip = net_generator.get_empty_string_sequence(bs)
+        text_features_flant5 = net_video_enc.get_empty_string_sequence(bs)
+        text_mask_flant5 = torch.zeros_like(text_features_flant5)
+        text_mask_flant5[:, 0] = 1
+
+    if negative_text is not None:
+        assert len(negative_text) == bs
+        negative_text_features_clip = feature_utils.encode_text_clip(negative_text)
+        negative_text_features_flant5, negative_text_mask_flant5 = feature_utils.encode_text_t5(negative_text)
+    else:
+        negative_text_features_clip = None
+        negative_text_features_flant5, negative_text_mask_flant5 = None, None
+    
+    if clip_video is not None:
+        clip_video = clip_video.to(device, dtype, non_blocking=True)
+        clip_features = feature_utils.encode_video_with_clip(clip_video,
+                                                             batch_size=bs *
+                                                             clip_batch_size_multiplier)
+        if image_input:
+            clip_features = clip_features.expand(-1, net_generator.clip_seq_len, -1)
+    else:
+        clip_features = net_generator.get_empty_clip_sequence(bs)
+
+    if sync_video is not None and not image_input:
+        text_features_flant5, text_mask_flant5 = net_video_enc.prepend_sup_text_tokens(text_features_flant5, text_mask_flant5)
+        sync_video = sync_video.to(net_video_enc.device, net_video_enc.dtype, non_blocking=True)
+        sync_features = net_video_enc.encode_video_with_sync(
+            sync_video, text_f=text_features_flant5, text_mask=text_mask_flant5
+        )
+    else:
+        sync_features = net_generator.get_empty_sync_sequence(bs)
+
+    x0 = torch.randn(bs,
+                     net_generator.latent_seq_len,
+                     net_generator.latent_dim,
+                     device=device,
+                     dtype=dtype,
+                     generator=rng)
+    preprocessed_conditions = net_generator.preprocess_conditions(clip_features, sync_features, text_features_clip)
+    empty_conditions = net_generator.get_empty_conditions(
+        bs, negative_text_features=negative_text_features_clip
+    )
+
+    cfg_ode_wrapper = lambda t, x: net_generator.ode_wrapper(t, x, preprocessed_conditions, empty_conditions,
+                                                   cfg_strength)
+    x1 = fm.to_data(cfg_ode_wrapper, x0)
+    x1 = net_generator.unnormalize(x1)
+    spec = feature_utils.decode(x1)
+    audio = feature_utils.vocode(spec)
+    return audio
+
+
+LOGFORMAT = "[%(log_color)s%(levelname)-8s%(reset)s]: %(log_color)s%(message)s%(reset)s"
+
+
+def setup_eval_logging(log_level: int = logging.INFO):
+    logging.root.setLevel(log_level)
+    formatter = ColoredFormatter(LOGFORMAT)
+    stream = logging.StreamHandler()
+    stream.setLevel(log_level)
+    stream.setFormatter(formatter)
+    log = logging.getLogger()
+    log.setLevel(log_level)
+    log.addHandler(stream)
+
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+def load_video(video_path: Path, duration_sec: float, load_all_frames: bool = True) -> VideoInfo:
+
+    clip_transform = v2.Compose([
+        v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+    ])
+
+    sync_transform = v2.Compose([
+        v2.Resize((_SYNC_SIZE, _SYNC_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+        # v2.CenterCrop(_SYNC_SIZE),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+        v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+    ])
+
+    output_frames, all_frames, orig_fps = read_frames(video_path,
+                                                      list_of_fps=[_CLIP_FPS, _SYNC_FPS],
+                                                      start_sec=0,
+                                                      end_sec=duration_sec,
+                                                      need_all_frames=load_all_frames)
+
+    clip_chunk, sync_chunk = output_frames
+    clip_chunk = torch.from_numpy(clip_chunk).permute(0, 3, 1, 2)
+    sync_chunk = torch.from_numpy(sync_chunk).permute(0, 3, 1, 2)
+
+    clip_frames = clip_transform(clip_chunk)
+    sync_frames = sync_transform(sync_chunk)
+
+    clip_length_sec = clip_frames.shape[0] / _CLIP_FPS
+    sync_length_sec = sync_frames.shape[0] / _SYNC_FPS
+
+    if clip_length_sec < duration_sec:
+        log.warning(f'Clip video is too short: {clip_length_sec:.2f} < {duration_sec:.2f}')
+        log.warning(f'Truncating to {clip_length_sec:.2f} sec')
+        duration_sec = clip_length_sec
+
+    if sync_length_sec < duration_sec:
+        log.warning(f'Sync video is too short: {sync_length_sec:.2f} < {duration_sec:.2f}')
+        log.warning(f'Truncating to {sync_length_sec:.2f} sec')
+        duration_sec = sync_length_sec
+
+    clip_frames = clip_frames[:int(_CLIP_FPS * duration_sec)]
+    sync_frames = sync_frames[:int(_SYNC_FPS * duration_sec)]
+
+    video_info = VideoInfo(
+        duration_sec=duration_sec,
+        fps=orig_fps,
+        clip_frames=clip_frames,
+        sync_frames=sync_frames,
+        all_frames=all_frames if load_all_frames else None,
+    )
+    return video_info
+
+
+def load_image(image_path: Path) -> VideoInfo:
+    clip_transform = v2.Compose([
+        v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+    ])
+
+    sync_transform = v2.Compose([
+        v2.Resize(_SYNC_SIZE, interpolation=v2.InterpolationMode.BICUBIC),
+        v2.CenterCrop(_SYNC_SIZE),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+        v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+    ])
+
+    frame = np.array(Image.open(image_path))
+
+    clip_chunk = torch.from_numpy(frame).unsqueeze(0).permute(0, 3, 1, 2)
+    sync_chunk = torch.from_numpy(frame).unsqueeze(0).permute(0, 3, 1, 2)
+
+    clip_frames = clip_transform(clip_chunk)
+    sync_frames = sync_transform(sync_chunk)
+
+    video_info = ImageInfo(
+        clip_frames=clip_frames,
+        sync_frames=sync_frames,
+        original_frame=frame,
+    )
+    return video_info
+
+
+def make_video(video_info: VideoInfo, output_path: Path, audio: torch.Tensor, sampling_rate: int):
+    reencode_with_audio(video_info, output_path, audio, sampling_rate)

selva_core/utils/log_integrator.py

@@ -0,0 +1,129 @@
+"""
+Integrate numerical values for some iterations
+Typically used for loss computation / logging to tensorboard
+Call finalize and create a new Integrator when you want to display/log
+"""
+from typing import Callable, Union
+
+import torch
+
+from selva_core.utils.logger import TensorboardLogger
+from selva_core.utils.tensor_utils import distribute_into_histogram
+
+
+class Integrator:
+
+    def __init__(self, logger: TensorboardLogger, distributed: bool = True):
+        self.values = {}
+        self.counts = {}
+        self.hooks = []  # List is used here to maintain insertion order
+
+        # for binned tensors
+        self.binned_tensors = {}
+        self.binned_tensor_indices = {}
+
+        self.logger = logger
+
+        self.distributed = distributed
+        self.local_rank = torch.distributed.get_rank()
+        self.world_size = torch.distributed.get_world_size()
+
+    def add_scalar(self, key: str, x: Union[torch.Tensor, int, float]):
+        if isinstance(x, torch.Tensor):
+            x = x.detach()
+            if x.dtype in [torch.long, torch.int, torch.bool]:
+                x = x.float()
+
+        if key not in self.values:
+            self.counts[key] = 1
+            self.values[key] = x
+        else:
+            self.counts[key] += 1
+            self.values[key] += x
+    
+    def add_scalar_with_count(self, key: str, x: Union[torch.Tensor, int, float], count: int):
+        if isinstance(x, torch.Tensor):
+            x = x.detach()
+            if x.dtype in [torch.long, torch.int, torch.bool]:
+                x = x.float()
+
+        if key not in self.values:
+            self.counts[key] = count
+            self.values[key] = x
+        else:
+            self.counts[key] += count
+            self.values[key] += x
+
+    def add_dict(self, tensor_dict: dict[str, torch.Tensor]):
+        for k, v in tensor_dict.items():
+            self.add_scalar(k, v)
+    
+    def add_dict_with_count(self, tensor_dict: dict[str, torch.Tensor], count: int):
+        for k, v in tensor_dict.items():
+            self.add_scalar_with_count(k, v, count)
+
+    def add_binned_tensor(self, key: str, x: torch.Tensor, indices: torch.Tensor):
+        if key not in self.binned_tensors:
+            self.binned_tensors[key] = [x.detach().flatten()]
+            self.binned_tensor_indices[key] = [indices.detach().flatten()]
+        else:
+            self.binned_tensors[key].append(x.detach().flatten())
+            self.binned_tensor_indices[key].append(indices.detach().flatten())
+
+    def add_hook(self, hook: Callable[[torch.Tensor], tuple[str, torch.Tensor]]):
+        """
+        Adds a custom hook, i.e. compute new metrics using values in the dict
+        The hook takes the dict as argument, and returns a (k, v) tuple
+        e.g. for computing IoU
+        """
+        self.hooks.append(hook)
+
+    def reset_except_hooks(self):
+        self.values = {}
+        self.counts = {}
+
+    # Average and output the metrics
+    def finalize(self, prefix: str, it: int, ignore_timer: bool = False) -> None:
+
+        for hook in self.hooks:
+            k, v = hook(self.values)
+            self.add_scalar(k, v)
+
+        # for the metrics
+        outputs = {}
+        for k, v in self.values.items():
+            avg = v / self.counts[k]
+            if self.distributed:
+                # Inplace operation
+                if isinstance(avg, torch.Tensor):
+                    avg = avg.cuda()
+                else:
+                    avg = torch.tensor(avg).cuda()
+                torch.distributed.reduce(avg, dst=0)
+
+                if self.local_rank == 0:
+                    avg = (avg / self.world_size).cpu().item()
+                    outputs[k] = avg
+            else:
+                # Simple does it
+                outputs[k] = avg
+
+        if (not self.distributed) or (self.local_rank == 0):
+            self.logger.log_metrics(prefix, outputs, it, ignore_timer=ignore_timer)
+
+        # for the binned tensors
+        for k, v in self.binned_tensors.items():
+            x = torch.cat(v, dim=0)
+            indices = torch.cat(self.binned_tensor_indices[k], dim=0)
+            hist, count = distribute_into_histogram(x, indices)
+
+            if self.distributed:
+                torch.distributed.reduce(hist, dst=0)
+                torch.distributed.reduce(count, dst=0)
+                if self.local_rank == 0:
+                    hist = hist / count
+            else:
+                hist = hist / count
+
+            if (not self.distributed) or (self.local_rank == 0):
+                self.logger.log_histogram(f'{prefix}/{k}', hist, it)

selva_core/utils/logger.py

@@ -0,0 +1,231 @@
+"""
+Dumps things to tensorboard and console
+"""
+
+import datetime
+import logging
+import math
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Optional, Union
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+import torchaudio
+from PIL import Image
+from pytz import timezone
+from torch.utils.tensorboard import SummaryWriter
+
+from selva_core.utils.email_utils import EmailSender
+from selva_core.utils.time_estimator import PartialTimeEstimator, TimeEstimator
+from selva_core.utils.timezone import my_timezone
+
+
+def tensor_to_numpy(image: torch.Tensor):
+    image_np = (image.numpy() * 255).astype('uint8')
+    return image_np
+
+
+def detach_to_cpu(x: torch.Tensor):
+    return x.detach().cpu()
+
+
+def fix_width_trunc(x: float):
+    return ('{:.9s}'.format('{:0.9f}'.format(x)))
+
+
+def plot_spectrogram(spectrogram: np.ndarray, title=None, ylabel="freq_bin", ax=None):
+    if ax is None:
+        _, ax = plt.subplots(1, 1)
+    if title is not None:
+        ax.set_title(title)
+    ax.set_ylabel(ylabel)
+    ax.imshow(spectrogram, origin="lower", aspect="auto", interpolation="nearest")
+
+
+class TensorboardLogger:
+
+    def __init__(self,
+                 exp_id: str,
+                 run_dir: Union[Path, str],
+                 py_logger: logging.Logger,
+                 *,
+                 is_rank0: bool = False,
+                 enable_email: bool = False):
+        self.exp_id = exp_id
+        self.run_dir = Path(run_dir)
+        self.py_log = py_logger
+        self.email_sender = EmailSender(exp_id, enable=(is_rank0 and enable_email))
+        if is_rank0:
+            self.tb_log = SummaryWriter(run_dir)
+        else:
+            self.tb_log = None
+
+        # Get current git info for logging
+        try:
+            import git
+            repo = git.Repo(".")
+            git_info = str(repo.active_branch) + ' ' + str(repo.head.commit.hexsha)
+        except (ImportError, RuntimeError, TypeError):
+            print('Failed to fetch git info. Defaulting to None')
+            git_info = 'None'
+
+        self.log_string('git', git_info)
+
+        # log the SLURM job id if available
+        job_id = os.environ.get('SLURM_JOB_ID', None)
+        if job_id is not None:
+            self.log_string('slurm_job_id', job_id)
+            self.email_sender.send(f'Job {job_id} started', f'Job started {run_dir}')
+
+        # used when logging metrics
+        self.batch_timer: TimeEstimator = None
+        self.data_timer: PartialTimeEstimator = None
+
+        self.nan_count = defaultdict(int)
+
+    def log_scalar(self, tag: str, x: float, it: int):
+        if self.tb_log is None:
+            return
+        if math.isnan(x) and 'grad_norm' not in tag:
+            self.nan_count[tag] += 1
+            if self.nan_count[tag] == 10:
+                self.email_sender.send(
+                    f'Nan detected in {tag} @ {self.run_dir}',
+                    f'Nan detected in {tag} at iteration {it}; run_dir: {self.run_dir}')
+        else:
+            self.nan_count[tag] = 0
+        self.tb_log.add_scalar(tag, x, it)
+
+    def log_metrics(self,
+                    prefix: str,
+                    metrics: dict[str, float],
+                    it: int,
+                    ignore_timer: bool = False):
+        msg = f'{self.exp_id}-{prefix} - it {it:6d}: '
+        metrics_msg = ''
+        for k, v in sorted(metrics.items()):
+            self.log_scalar(f'{prefix}/{k}', v, it)
+            metrics_msg += f'{k: >10}:{v:.7f},\t'
+
+        if self.batch_timer is not None and not ignore_timer:
+            self.batch_timer.update()
+            avg_time = self.batch_timer.get_and_reset_avg_time()
+            data_time = self.data_timer.get_and_reset_avg_time()
+
+            # add time to tensorboard
+            self.log_scalar(f'{prefix}/avg_time', avg_time, it)
+            self.log_scalar(f'{prefix}/data_time', data_time, it)
+
+            est = self.batch_timer.get_est_remaining(it)
+            est = datetime.timedelta(seconds=est)
+            if est.days > 0:
+                remaining_str = f'{est.days}d {est.seconds // 3600}h'
+            else:
+                remaining_str = f'{est.seconds // 3600}h {(est.seconds%3600) // 60}m'
+            eta = datetime.datetime.now(timezone(my_timezone)) + est
+            eta_str = eta.strftime('%Y-%m-%d %H:%M:%S %Z%z')
+            time_msg = f'avg_time:{avg_time:.3f},data:{data_time:.3f},remaining:{remaining_str},eta:{eta_str},\t'
+            msg = f'{msg} {time_msg}'
+
+        msg = f'{msg} {metrics_msg}'
+        self.py_log.info(msg)
+
+    def log_histogram(self, tag: str, hist: torch.Tensor, it: int):
+        if self.tb_log is None:
+            return
+        # hist should be a 1D tensor
+        hist = hist.cpu().numpy()
+        fig, ax = plt.subplots()
+        x_range = np.linspace(0, 1, len(hist))
+        ax.bar(x_range, hist, width=1 / (len(hist) - 1))
+        ax.set_xticks(x_range)
+        ax.set_xticklabels(x_range)
+        plt.tight_layout()
+        self.tb_log.add_figure(tag, fig, it)
+        plt.close()
+
+    def log_image(self, prefix: str, tag: str, image: np.ndarray, it: int):
+        image_dir = self.run_dir / f'{prefix}_images'
+        image_dir.mkdir(exist_ok=True, parents=True)
+
+        image = Image.fromarray(image)
+        image.save(image_dir / f'{it:09d}_{tag}.png')
+
+    def log_audio(self,
+                  prefix: str,
+                  tag: str,
+                  waveform: torch.Tensor,
+                  it: Optional[int] = None,
+                  *,
+                  subdir: Optional[Path] = None,
+                  sample_rate: int = 16000) -> Path:
+        if subdir is None:
+            audio_dir = self.run_dir / prefix
+        else:
+            audio_dir = self.run_dir / subdir / prefix
+        audio_dir.mkdir(exist_ok=True, parents=True)
+
+        if it is None:
+            name = f'{tag}.flac'
+        else:
+            name = f'{it:09d}_{tag}.flac'
+
+        torchaudio.save(audio_dir / name,
+                        waveform.cpu().float(),
+                        sample_rate=sample_rate,
+                        channels_first=True)
+        return Path(audio_dir)
+
+    def log_spectrogram(
+        self,
+        prefix: str,
+        tag: str,
+        spec: torch.Tensor,
+        it: Optional[int],
+        *,
+        subdir: Optional[Path] = None,
+    ):
+        if subdir is None:
+            spec_dir = self.run_dir / prefix
+        else:
+            spec_dir = self.run_dir / subdir / prefix
+        spec_dir.mkdir(exist_ok=True, parents=True)
+
+        if it is None:
+            name = f'{tag}.png'
+        else:
+            name = f'{it:09d}_{tag}.png'
+
+        plot_spectrogram(spec.cpu().float())
+        plt.tight_layout()
+        plt.savefig(spec_dir / name)
+        plt.close()
+
+    def log_string(self, tag: str, x: str):
+        self.py_log.info(f'{tag} - {x}')
+        if self.tb_log is None:
+            return
+        self.tb_log.add_text(tag, x)
+
+    def debug(self, x):
+        self.py_log.debug(x)
+
+    def info(self, x):
+        self.py_log.info(x)
+
+    def warning(self, x):
+        self.py_log.warning(x)
+
+    def error(self, x):
+        self.py_log.error(x)
+
+    def critical(self, x):
+        self.py_log.critical(x)
+
+        self.email_sender.send(f'Error occurred in {self.run_dir}', x)
+
+    def complete(self):
+        self.email_sender.send(f'Job completed in {self.run_dir}', 'Job completed')

selva_core/utils/misc.py

@@ -0,0 +1,41 @@
+import importlib
+from typing import Optional
+
+from omegaconf import DictConfig, OmegaConf
+
+
+
+def instantiate_from_config(target: str, params: Optional[dict] = None):
+    """
+    Instantiate an object from a dotted path `target` and keyword `params`.
+    Common name: instantiate_from_config
+    """
+    if not target or not isinstance(target, str):
+        raise ValueError(f"Invalid target: {target!r}")
+    params = {} if params is None else params
+
+    # Convert OmegaConf DictConfig to plain dict if needed
+    try:
+        if isinstance(params, DictConfig):
+            params = OmegaConf.to_container(params, resolve=True)
+    except Exception:
+        pass
+
+    try:
+        module_path, attr_name = target.rsplit('.', 1)
+    except ValueError as e:
+        raise ValueError(f"Target must be like 'pkg.mod.Class', got {target!r}") from e
+
+    try:
+        module = importlib.import_module(module_path)
+    except ModuleNotFoundError as e:
+        raise ModuleNotFoundError(f"Could not import module '{module_path}' for target '{target}'.") from e
+
+    try:
+        obj = getattr(module, attr_name)
+    except AttributeError as e:
+        raise AttributeError(f"Module '{module_path}' has no attribute '{attr_name}' (from '{target}').") from e
+
+    if not callable(obj):
+        raise TypeError(f"Resolved target '{target}' is not callable.")
+    return obj(**dict(params))

selva_core/utils/synthesize_ema.py

@@ -0,0 +1,22 @@
+from typing import Optional
+
+from nitrous_ema import PostHocEMA
+from omegaconf import DictConfig
+
+from selva_core.model.utils.factory import create_model_from_factory
+
+
+def synthesize_ema(cfg: DictConfig, sigma: float, step: Optional[int]):
+    vae = create_model_from_factory(cfg.model.factory_path,
+                                    cfg.model.name,
+                                    **cfg.model.get('params', {})
+                                    )
+    emas = PostHocEMA(vae,
+                      sigma_rels=cfg.ema.sigma_rels,
+                      update_every=cfg.ema.update_every,
+                      checkpoint_every_num_steps=cfg.ema.checkpoint_every,
+                      checkpoint_folder=cfg.ema.checkpoint_folder)
+
+    synthesized_ema = emas.synthesize_ema_model(sigma_rel=sigma, step=step, device='cpu')
+    state_dict = synthesized_ema.ema_model.state_dict()
+    return state_dict