chore: sanitize tooltips/comments + add experiment configs

- Replace all BJ references with generic "target style/audio" in activation steering, DITTO optimizer, and BigVGAN trainer - Add latent_mixup_alpha/latent_noise_sigma to LoRA scheduler defaults - Add bigvgan_disc_fm_retest.json and lora_optimized_dataset.json Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-04-10 13:44:37 +02:00
parent 082a2da438
commit f745e241c4
6 changed files with 120 additions and 23 deletions
@@ -0,0 +1,31 @@
+{
+  "name": "bigvgan_disc_fm_retest",
+  "description": "Retest discriminator feature matching after bfloat16 dtype fix. Uses optimal config from overnight sweep (snake_alpha, GAFilter, lr=1e-4, phase=1.0, L2-SP=1e-3, 5000 steps).",
+  "data_dir": "/media/unraid/davinci/Selva/BJ/features",
+  "output_root": "/media/unraid/davinci/Selva/BJ/experiment/bigvgan_disc_fm_retest",
+  "base": {
+    "train_mode": "snake_alpha_only",
+    "steps": 5000,
+    "lr": 1e-4,
+    "batch_size": 8,
+    "segment_seconds": 0.5,
+    "lambda_l2sp": 1e-3,
+    "use_gafilter": true,
+    "gafilter_kernel_size": 9,
+    "lambda_phase": 1.0,
+    "save_every": 1000,
+    "seed": 42,
+    "lora_adapter": "/media/unraid/davinci/Selva/BJ/experiment/pissa_sweep/standard_baseline/adapter_final.pt"
+  },
+  "experiments": [
+    {
+      "id": "snake_5k_control",
+      "description": "Control: best config from overnight sweep without discriminator. Baseline for A/B comparison."
+    },
+    {
+      "id": "disc_fm_5k",
+      "description": "Discriminator feature matching at 5k steps. Tests if perceptual FM loss improves over mel+phase alone.",
+      "discriminator_path": "/media/unraid/davinci/Selva/BJ/experiment/bigvgan_discriminator_optimizer.pt"
+    }
+  ]
+}
@@ -0,0 +1,64 @@
+{
+  "name": "lora_optimized_dataset",
+  "description": "LoRA training on optimized dataset (134 clips: resampled 44.1kHz, LUFS-normalized, spectral matched, HF smoothed, gain-augmented). Tests latent augmentation and schedule variants on top of known-best config (PiSSA, rank=128, lr=3e-4).",
+  "data_dir": "/media/unraid/davinci/Selva/BJ/features_v2_improved/",
+  "output_root": "/media/unraid/davinci/Selva/BJ/experiment/lora_optimized_dataset",
+  "base": {
+    "rank": 128,
+    "lr": 3e-4,
+    "steps": 5000,
+    "batch_size": 4,
+    "warmup_steps": 100,
+    "save_every": 1000,
+    "seed": 42,
+    "init_mode": "pissa",
+    "use_rslora": true,
+    "target": "attn.qkv",
+    "timestep_mode": "uniform",
+    "lr_schedule": "constant"
+  },
+  "experiments": [
+    {
+      "id": "baseline",
+      "description": "Control: known-best config (PiSSA r128 lr=3e-4) on the optimized dataset. No latent augmentation."
+    },
+    {
+      "id": "latent_mixup",
+      "description": "Latent mixup alpha=0.4 (MusicLDM). Tests if mixing training latents reduces memorization on 134 clips.",
+      "latent_mixup_alpha": 0.4
+    },
+    {
+      "id": "latent_noise",
+      "description": "Latent noise sigma=0.02. Mild Gaussian noise on training latents for regularization.",
+      "latent_noise_sigma": 0.02
+    },
+    {
+      "id": "mixup_and_noise",
+      "description": "Both latent mixup (0.4) and noise (0.02). Combined regularization.",
+      "latent_mixup_alpha": 0.4,
+      "latent_noise_sigma": 0.02
+    },
+    {
+      "id": "cosine_schedule",
+      "description": "Cosine LR decay. lr=3e-4 was stable with constant, but cosine may extract more from 5k steps.",
+      "lr_schedule": "cosine"
+    },
+    {
+      "id": "cosine_mixup",
+      "description": "Cosine LR + latent mixup. Best regularization combo candidate.",
+      "lr_schedule": "cosine",
+      "latent_mixup_alpha": 0.4
+    },
+    {
+      "id": "logit_normal",
+      "description": "Logit-normal timestep sampling (sigma=1.0). Concentrates training near t=0.5 where flow matching is hardest.",
+      "timestep_mode": "logit_normal"
+    },
+    {
+      "id": "curriculum_mixup",
+      "description": "Curriculum timesteps (logit_normal first 60%, then uniform) + latent mixup. Full regularization stack.",
+      "timestep_mode": "curriculum",
+      "latent_mixup_alpha": 0.4
+    }
+  ]
+}