36 lines
1.4 KiB
JSON
36 lines
1.4 KiB
JSON
{
|
|
"name": "bigvgan_optimized_dataset",
|
|
"description": "BigVGAN fine-tuning on optimized dataset (134 clips, 44.1kHz, LUFS-normalized). Standard mode (no LoRA) — trains decoder to faithfully reconstruct target domain audio from mel spectrograms. Uses optimal config from prior sweeps.",
|
|
"data_dir": "/media/unraid/davinci/Selva/BJ/features_v2_improved/",
|
|
"output_root": "/media/unraid/davinci/Selva/BJ/experiment/bigvgan_optimized_dataset",
|
|
"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
|
|
},
|
|
"experiments": [
|
|
{
|
|
"id": "standard_5k",
|
|
"description": "Standard mode: mel from clean FLAC → BigVGAN → reconstruct FLAC. No LoRA. Directly improves VAE roundtrip quality."
|
|
},
|
|
{
|
|
"id": "disc_fm_5k",
|
|
"description": "Standard mode + discriminator feature matching. Tests if perceptual loss helps on clean audio reconstruction.",
|
|
"discriminator_path": "/media/unraid/davinci/Selva/BJ/experiment/bigvgan_discriminator_optimizer.pt"
|
|
},
|
|
{
|
|
"id": "standard_10k",
|
|
"description": "Extended 10k steps. More data passes on 134 clips may extract more from the optimized dataset.",
|
|
"steps": 10000
|
|
}
|
|
]
|
|
}
|