Diff-VocSep/config.yaml at main · chynggi/Diff-VocSep · 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
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
# 기본 설정 값 (필요 시 수정)
seed: 42

# 데이터
data:
  musdb_root: ./data/musdb18
  musdbhq_root: ./data/musdb18hq
  segment_seconds: 2.0
  mono: true
  dataset: musdbhq  # musdb | musdbhq
  pitch_aug:
    enabled: false
    semitones: [-2.0, 2.0]
    prob: 0.3

# 오디오 / STFT
audio:
  # MSST 스타일 기본값 (필요 시 조정)
  sample_rate: 44100
  n_fft: 8192
  hop_length: 1024
  win_length: 8192
  window: hann
  center: true

# 학습
train:
  batch_size: 1
  epochs: 100
  lr: 2.0e-4
  weight_decay: 1.0e-2
  grad_clip: 1.0
  amp: true
  log_interval: 50
  val_every_steps: 75
  val_batches: 8
  loss_cf_weight: 1.0
  loss_voc_weight: 0.5

# 확산
diffusion:
  timesteps: 200
  beta_start: 1.0e-4
  beta_end: 2.0e-2
  beta_schedule: linear  # linear | cosine
  use_ddim: true
  ddim_steps: 50
  val_ddim_steps: 50
  eta: 0.0
  # sampler: ddpm | ddim | (future: dpm-solver, unipc)
  sampler: ddim
  # Validation shallow diffusion options
  validate_use_shallow: true
  shallow_k: 50
  add_forward_noise: true

# 모델
model:
  in_channels: 2   # [noisy, mixture]
  out_channels: 1  # instruments magnitude
  base_channels: 64
  channels_mult: [1, 2, 4]
  attn_resolutions: [4, 8]
  model_type: conformer  # unet | conformer
  model_kwargs:
    d_model: 128
    n_heads: 4
    d_ff: 256
    num_layers: 4
    kernel_size: 31
    dropout: 0.0
    axis: time  # time | freq | mixed
    # Memory helpers (safe defaults)
    freq_chunk: 64   # process freq bins in chunks when axis='time' to reduce memory
    time_chunk: null # set e.g., 128 if axis='freq'
    use_checkpoint: false

# 로깅/체크포인트
log:
  out_dir: ./logs
  use_wandb: false
  use_tensorboard: true
  tb_log_dir: ./logs/tb
  project: diff-msst
  entity: null
  save_best: true
  save_last: true