forked from jinhan/tacotron2-vae
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathhparams.yaml
110 lines (95 loc) · 2.95 KB
/
hparams.yaml
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
################################
# Experiment Parameters #
################################
epochs: 300
iters_per_checkpoint: 500
seed: 42
dynamic_loss_scaling: True
fp16_run: False
distributed_run: False
dist_backend: "nccl"
dist_url: "tcp://localhost:54321"
cudnn_enabled: True
cudnn_benchmark: True
################################
# Data Parameters #
################################
load_mel_from_disk: False
training_files: 'filelists/arthur_v3_train.txt'
validation_files: 'filelists/arthur_v3_valid.txt'
text_cleaners: ["english_cleaners"] # english_cleaners, korean_cleaners
sort_by_length: False
################################
# Audio Parameters #
################################
max_wav_value: 32768.0 # NOTE, all audio parameters the same as HiFiGAN vocoder v1 except sample rate (22050)
sampling_rate: 16000
filter_length: 1024
hop_length: 256 # number audio of frames between stft colmns, default win_length/4
win_length: 1024 # win_length int <= n_ftt: fft window size (frequency domain), defaults to win_length: n_fft
n_mel_channels: 80
mel_fmin: 0.0
mel_fmax: 8000.0
################################
# Model Parameters #
################################
n_symbols: 80 # set 80 if u use korean_cleaners. set 65 if u use english_cleaners
symbols_embedding_dim: 512
# Transcript encoder parameters
encoder_kernel_size: 5
encoder_n_convolutions: 3
encoder_embedding_dim: 512
# Speaker embedding parameters
n_speakers: 1
speaker_embedding_dim: 16
# ---------------------------------------- #
# emotion
n_emotions: 7 # number of emotion labels
emotion_embedding_dim: 16
# reference encoder
E: 512
ref_enc_filters: [32, 32, 64, 64, 128, 128]
ref_enc_size: [3, 3]
ref_enc_strides: [2, 2]
ref_enc_pad: [1, 1]
ref_enc_gru_size: 256 # was 512 // 2
z_latent_dim: 16
anneal_function: logistic # 'constant' is simple beta-vae
anneal_k: 0.0025
anneal_x0: 5000 # Previously 10000. This number chosen after some plotting
anneal_upper: 0.2
anneal_lag: 50000
# Prosody embedding parameters
prosody_n_convolutions: 6
prosody_conv_dim_in: [1, 32, 32, 64, 64, 128]
prosody_conv_dim_out: [32, 32, 64, 64, 128, 128]
prosody_conv_kernel: 3
prosody_conv_stride: 2
prosody_embedding_dim: 128
# Decoder parameters
n_frames_per_step: 1 # currently only 1 is supported
decoder_rnn_dim: 1024
prenet_dim: 256
max_decoder_steps: 1000
gate_threshold: 0.5
p_attention_dropout: 0.1
p_decoder_dropout: 0.1
# Attention parameters
attention_rnn_dim: 1024
attention_dim: 128
# Location Layer parameters
attention_location_n_filters: 32
attention_location_kernel_size: 31
# Mel-post processing network parameters
postnet_embedding_dim: 512
postnet_kernel_size: 5
postnet_n_convolutions: 5
################################
# Optimization Hyperparameters #
################################
use_saved_learning_rate: False
learning_rate: !!float 1e-3
weight_decay: !!float 1e-6
grad_clip_thresh: 1.0
batch_size: 16
mask_padding: True # set model's padded outputs to padded values