waveglow-vqvae/config.yml at master · jaywalnut310/waveglow-vqvae · 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
# audio
sample_rate: 22050
mel_channels: 80
fmin: 0.0
fmax: 8000.0
fft_size: 1024
hop_size: 256
max_wav_value: 32768.0

# model
n_flows: 12
n_group: 8
n_early_every: 4
n_early_size: 2
n_layers: 8
n_channels: 256
kernel_size: 3
sigma: 1.0
act_name: leaky_relu # leaky_relu, relu, relu6

# conditon
freq_axis_kernel_size: 3
upsample_scales: [8, 8]
use_film: true
use_cond_wn: false

# speaker ids
emb_channels: 512
n_speakers: 512

# vq
use_vq: true
bottleneck_bits: 9
bottleneck_dims_per_bit: 16
hidden_channels: 256
beta: 0.01
gamma: 0.01

# training
ftype: float16 # float16, float32
local_condition: wav # mel, wav
global_condition: null # null, sid
max_input_length: 16384
train_steps: 1000000
eval_steps: null
throttle_secs: 10
save_summary_steps: 1000
save_checkpoints_secs: 1800
n_checkpoints: 5
train_batch_size: 6
learning_rate: 0.0001
min_lr: 0.00005
lr_decay: halve # null, noam, halve
halve_step: 500000
exponential_moving_average: false
ema_decay: 0.9999

# optimizer
optimizer: adam
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 0.00000001
clip_gradients: null
loss_scale: 512 # 1, 512

# eval
eval_batch_size: 1

# infer
infer_batch_size: 1

# data paths
vocab_file: null
load_features: [wav, mel]
tfr_dir: datasets/ljs
tfr_prefix: ljs
train_files: filelists/ljs_audio_text_train_filelist.txt
eval_files:  filelists/ljs_audio_text_eval_filelist.txt