feat: add dataset pipeline nodes + latent augmentation for LoRA trainer

New dataset pipeline nodes:
- SelvaDatasetSpectralMatcher: batch spectral EQ toward VAE distribution
- SelvaDatasetHfSmoother: batch HF attenuation for codec compatibility
- SelvaDatasetAugmenter: gain/pitch/time-stretch variants with npz origin tracking

Improvements:
- Inspector: silence detection (max_silence_fraction param)
- Saver: origin_name lookup for augmented clips' npz pairing
- LoRA trainer: latent_mixup_alpha + latent_noise_sigma regularization
- LoRA trainer: one-time SR mismatch warning in _load_audio

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-04-10 11:32:00 +02:00
parent 30127c13ca
commit af6c225f53
3 changed files with 299 additions and 3 deletions
+3
View File
@@ -32,6 +32,9 @@ _NODES = {
"SelvaDatasetSaver": (".selva_dataset_pipeline", "SelvaDatasetSaver", "SelVA Dataset Saver"),
"SelvaHarmonicExciter": (".selva_audio_postprocess", "SelvaHarmonicExciter", "SelVA Harmonic Exciter"),
"SelvaOutputNormalizer": (".selva_audio_postprocess", "SelvaOutputNormalizer", "SelVA Output Normalizer"),
"SelvaDatasetSpectralMatcher": (".selva_dataset_pipeline", "SelvaDatasetSpectralMatcher", "SelVA Dataset Spectral Matcher"),
"SelvaDatasetHfSmoother": (".selva_dataset_pipeline", "SelvaDatasetHfSmoother", "SelVA Dataset HF Smoother"),
"SelvaDatasetAugmenter": (".selva_dataset_pipeline", "SelvaDatasetAugmenter", "SelVA Dataset Augmenter"),
}
for key, (module_path, class_name, display_name) in _NODES.items():