final_models/audio_spectrogram_vit/config.json

{
  "cache_dir": "training/audio_spectrogram_vit/feature_cache",
  "display_name": "SSLAM ViT-Base audio spectrogram",
  "feature_config": {
    "feature": "SSLAM extract_features embedding",
    "feature_shape": [
      768
    ],
    "head": "MLP logit over cached SSLAM embedding",
    "mel_bins": 128,
    "model_id": "ta012/SSLAM_pretrain",
    "norm_mean": -4.268,
    "norm_std": 4.569,
    "sample_rate": 16000,
    "target_frames": 1024
  },
  "frozen_backbone": "ta012/SSLAM_pretrain",
  "method": "audio_spectrogram_vit",
  "train_config": {
    "batch_size": 8,
    "cache_root": "training",
    "dataset_root": "dataset",
    "device": "auto",
    "epochs": 20,
    "learning_rate": 0.0003,
    "num_workers": 0,
    "output_root": "final_models",
    "patience": 5,
    "print_progress": true,
    "seed": 153,
    "use_amp": true,
    "weight_decay": 0.0001
  }
}