Hugging Face's logo

Spaces:
facebook
/
MusicGen
Running on A10G

App Files Community
78
MusicGen
File size: 2,955 Bytes 
5325fcc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
 
# @package __global__

defaults:
  - ../default
  - override /dset: audio/default
  - _self_

solver: compression
sample_rate: ???
channels: ???

# loss balancing
losses:
  adv: 4.
  feat: 4.
  l1: 0.1
  mel: 0.
  msspec: 2.
  sisnr: 0.
balancer:
  balance_grads: true
  ema_decay: 0.999
  per_batch_item: true
  total_norm: 1.

adversarial:
  every: 1
  adversaries: [msstftd]
  adv_loss: hinge
  feat_loss: l1

# losses hyperparameters
l1: {}
l2: {}
mrstft:
  factor_sc: .5
  factor_mag: .5
  normalized: false
mel:
  sample_rate: ${sample_rate}
  n_fft: 1024
  hop_length: 256
  win_length: 1024
  n_mels: 64
  f_min: 64
  f_max: null
  normalized: false
  floor_level: 1e-5
sisnr:
  sample_rate: ${sample_rate}
  segment: 5.
msspec:
  sample_rate: ${sample_rate}
  range_start: 6
  range_end: 11
  n_mels: 64
  f_min: 64
  f_max: null
  normalized: true
  alphas: false
  floor_level: 1e-5

# metrics
metrics:
  visqol:
    mode: audio
    bin: null  # path to visqol install
    model: tcdaudio14_aacvopus_coresv_svrnsim_n.68_g.01_c1.model # visqol v3

# adversaries hyperparameters
msstftd:
  in_channels: 1
  out_channels: 1
  filters: 32
  norm: weight_norm
  n_ffts: [1024, 2048, 512, 256, 128]
  hop_lengths: [256, 512, 128, 64, 32]
  win_lengths: [1024, 2048, 512, 256, 128]
  activation: LeakyReLU
  activation_params: {negative_slope: 0.3}
msd:
  in_channels: 1
  out_channels: 1
  scale_norms: [spectral_norm, weight_norm, weight_norm]
  kernel_sizes: [5, 3]
  filters: 16
  max_filters: 1024
  downsample_scales: [4, 4, 4, 4]
  inner_kernel_sizes: null
  groups: [4, 4, 4, 4]
  strides: null
  paddings: null
  activation: LeakyReLU
  activation_params: {negative_slope: 0.3}
mpd:
  in_channels: 1
  out_channels: 1
  periods: [2, 3, 5, 7, 11]
  n_layers: 5
  kernel_size: 5
  stride: 3
  filters: 8
  filter_scales: 4
  max_filters: 1024
  activation: LeakyReLU
  activation_params: {negative_slope: 0.3}
  norm: weight_norm

# data hyperparameters
dataset:
  batch_size: 64
  num_workers: 10
  segment_duration: 1
  train:
    num_samples: 500000
  valid:
    num_samples: 10000
  evaluate:
    batch_size: 32
    num_samples: 10000
  generate:
    batch_size: 32
    num_samples: 50
    segment_duration: 10

# solver hyperparameters
evaluate:
  every: 25
  num_workers: 5
  metrics:
    visqol: false
    sisnr: true
generate:
  every: 25
  num_workers: 5
  audio:
    sample_rate: ${sample_rate}

# checkpointing schedule
checkpoint:
  save_last: true
  save_every: 25
  keep_last: 10
  keep_every_states: null

# optimization hyperparameters
optim:
  epochs: 200
  updates_per_epoch: 2000
  lr: 3e-4
  max_norm: 0.
  optimizer: adam
  adam:
    betas: [0.5, 0.9]
    weight_decay: 0.
  ema:
    use: true         # whether to use EMA or not
    updates: 1        # update at every step
    device: ${device} # device for EMA, can be put on GPU if more frequent updates
    decay: 0.99       # EMA decay value, if null, no EMA is used