LIVI-Lyrics-Informed-Version-Identification/src/livi/apps/audio_encoder/config/livi.yaml at main · deezer/LIVI-Lyrics-Informed-Version-Identification · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
# ============================== GLOBAL ==============================
seed: 1234

# ============================== DATA ================================
data:
  data_dir:            /nfs/interns/jaffolter/data/audio_encoder_dataset
  last_shard_train:    "001339"     # last shard for training set
  last_shard_val:      "000167"     # last shard for validation set
  last_shard_test:     "000002"     # last shard for test set
  window:              1000         # window size for WebDataset
  sr:                  16000        # sampling rate (16k for Whisper)
  num_workers:         8
  prefetch_factor:     4
  batch_size:          128
  total_train_samples: 1339790
  total_val_samples:   167484
  model_dir:           livi/apps/audio_encoder/models

# ============================== MODEL ================================
model:
  _target_:        livi.apps.audio_encoder.models.LiviAudioEncoder
  checkpoint_dir:  src/livi/apps/audio_encoder/checkpoints
  device:          cuda:0

  # Whisper encoder
  dim_whisper:         1280
  whisper_model_name: "openai/whisper-large-v3-turbo"
  compile:             false

  # Attention Pooling
  num_heads:       1
  mlp_ratio:       2.0
  qkv_bias:        false
  qk_scale:        null
  drop:            0.0
  attn_drop:       0.0
  init_scale:      1e-4

  # Projection head (audio)
  dim_embed:       768
  dim_hiddens:     [3072, 2048, 2048, 1536]

# ============================== OPTIMIZER ============================
optimizer:
  lr:              1e-4
  max_lr:          5e-4
  weight_decay:    0.01
  opt_betas:       [0.9, 0.98]
  warmup_steps:    10000
  alpha_loss:      0.5

# ============================== TRAINING =============================
train:
  epochs:             3

# ============================== LOGGING ==============================
wandb:
  project:          LIVI
  log_dir:          src/livi/apps/audio_encoder/logs
  log_steps:        200

# ============================== METRICS ==============================
metrics:
  topk: [1, 5, 10]