File size: 8,292 Bytes
27ddcef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb57397
 
27ddcef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eb57397
27ddcef
eb57397
27ddcef
eb57397
27ddcef
 
 
 
eb57397
27ddcef
eb57397
27ddcef
 
 
 
 
 
 
 
 
 
 
 
 
 
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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
# This configuration is for ESPnet2 to train VITS, which
# is truely end-to-end text-to-waveform model. To run
# this config, you need to specify "--tts_task gan_tts"
# option for tts.sh at least and use 22050 hz audio as
# the training data (mainly tested on LJspeech).
# This configuration tested on 4 GPUs (V100) with 32GB GPU
# memory. It takes around 2 weeks to finish the training
# but 100k iters model should generate reasonable results.

##########################################################
#                  TTS MODEL SETTING                     #
##########################################################
tts: vits
tts_conf:
    # generator related
    generator_type: vits_generator
    generator_params:
        hidden_channels: 192
        spks: 128
        global_channels: 256
        segment_size: 32
        text_encoder_attention_heads: 2
        text_encoder_ffn_expand: 4
        text_encoder_blocks: 6
        text_encoder_positionwise_layer_type: "conv1d"
        text_encoder_positionwise_conv_kernel_size: 3
        text_encoder_positional_encoding_layer_type: "rel_pos"
        text_encoder_self_attention_layer_type: "rel_selfattn"
        text_encoder_activation_type: "swish"
        text_encoder_normalize_before: true
        text_encoder_dropout_rate: 0.1
        text_encoder_positional_dropout_rate: 0.0
        text_encoder_attention_dropout_rate: 0.1
        use_macaron_style_in_text_encoder: true
        # NOTE(kan-bayashi): Conformer conv requires BatchNorm1d which causes
        #   errors when multiple GPUs in pytorch 1.7.1. Therefore, we disable
        #   it as a default. We need to consider the alternative normalization
        #   or different version pytorch may solve this issue.
        use_conformer_conv_in_text_encoder: false
        text_encoder_conformer_kernel_size: -1
        decoder_kernel_size: 7
        decoder_channels: 512
        decoder_upsample_scales: [8, 8, 2, 2]
        decoder_upsample_kernel_sizes: [16, 16, 4, 4]
        decoder_resblock_kernel_sizes: [3, 7, 11]
        decoder_resblock_dilations: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
        use_weight_norm_in_decoder: true
        posterior_encoder_kernel_size: 5
        posterior_encoder_layers: 16
        posterior_encoder_stacks: 1
        posterior_encoder_base_dilation: 1
        posterior_encoder_dropout_rate: 0.0
        use_weight_norm_in_posterior_encoder: true
        flow_flows: 4
        flow_kernel_size: 5
        flow_base_dilation: 1
        flow_layers: 4
        flow_dropout_rate: 0.0
        use_weight_norm_in_flow: true
        use_only_mean_in_flow: true
        stochastic_duration_predictor_kernel_size: 3
        stochastic_duration_predictor_dropout_rate: 0.5
        stochastic_duration_predictor_flows: 4
        stochastic_duration_predictor_dds_conv_layers: 3
    # discriminator related
    discriminator_type: hifigan_multi_scale_multi_period_discriminator
    discriminator_params:
        scales: 1
        scale_downsample_pooling: "AvgPool1d"
        scale_downsample_pooling_params:
            kernel_size: 4
            stride: 2
            padding: 2
        scale_discriminator_params:
            in_channels: 1
            out_channels: 1
            kernel_sizes: [15, 41, 5, 3]
            channels: 128
            max_downsample_channels: 1024
            max_groups: 16
            bias: True
            downsample_scales: [2, 2, 4, 4, 1]
            nonlinear_activation: "LeakyReLU"
            nonlinear_activation_params:
                negative_slope: 0.1
            use_weight_norm: True
            use_spectral_norm: False
        follow_official_norm: False
        periods: [2, 3, 5, 7, 11]
        period_discriminator_params:
            in_channels: 1
            out_channels: 1
            kernel_sizes: [5, 3]
            channels: 32
            downsample_scales: [3, 3, 3, 3, 1]
            max_downsample_channels: 1024
            bias: True
            nonlinear_activation: "LeakyReLU"
            nonlinear_activation_params:
                negative_slope: 0.1
            use_weight_norm: True
            use_spectral_norm: False
    # loss function related
    generator_adv_loss_params:
        average_by_discriminators: false # whether to average loss value by #discriminators
        loss_type: mse                   # loss type, "mse" or "hinge"
    discriminator_adv_loss_params:
        average_by_discriminators: false # whether to average loss value by #discriminators
        loss_type: mse                   # loss type, "mse" or "hinge"
    feat_match_loss_params:
        average_by_discriminators: false # whether to average loss value by #discriminators
        average_by_layers: false         # whether to average loss value by #layers of each discriminator
        include_final_outputs: true      # whether to include final outputs for loss calculation
    mel_loss_params:
        fs: 22050          # must be the same as the training data
        n_fft: 1024        # fft points
        hop_length: 256    # hop size
        win_length: null   # window length
        window: hann       # window type
        n_mels: 80         # number of Mel basis
        fmin: 0            # minimum frequency for Mel basis
        fmax: null         # maximum frequency for Mel basis
        log_base: null     # null represent natural log
    lambda_adv: 1.0        # loss scaling coefficient for adversarial loss
    lambda_mel: 45.0       # loss scaling coefficient for Mel loss
    lambda_feat_match: 2.0 # loss scaling coefficient for feat match loss
    lambda_dur: 1.0        # loss scaling coefficient for duration loss
    lambda_kl: 1.0         # loss scaling coefficient for KL divergence loss
    # others
    sampling_rate: 22050          # needed in the inference for saving wav
    cache_generator_outputs: true # whether to cache generator outputs in the training

##########################################################
#            OPTIMIZER & SCHEDULER SETTING               #
##########################################################
# optimizer setting for generator
optim: adamw
optim_conf:
    lr: 2.0e-4
    betas: [0.8, 0.99]
    eps: 1.0e-9
    weight_decay: 0.0
scheduler: exponentiallr
scheduler_conf:
    gamma: 0.999875
# optimizer setting for discriminator
optim2: adamw
optim2_conf:
    lr: 2.0e-4
    betas: [0.8, 0.99]
    eps: 1.0e-9
    weight_decay: 0.0
scheduler2: exponentiallr
scheduler2_conf:
    gamma: 0.999875
generator_first: false # whether to start updating generator first

##########################################################
#                OTHER TRAINING SETTING                  #
##########################################################
#num_iters_per_epoch: 1000 # number of iterations per epoch
max_epoch: 30           # number of epochs
accum_grad: 1             # gradient accumulation
batch_bins: 1900000       # batch bins (feats_type=raw)
batch_type: numel         # how to make batch
#batch_type: sorted           # how to make batchbatch_size: 1
grad_clip: -1             # gradient clipping norm
grad_noise: false         # whether to use gradient noise injection
sort_in_batch: descending # how to sort data in making batch
sort_batch: descending    # how to sort created batches
num_workers: 1            # number of workers of data loader
use_amp: false            # whether to use pytorch amp
train_dtype: float32
log_interval: 50          # log interval in iterations
keep_nbest_models: 10     # number of models to keep
num_att_plot: 3           # number of attention figures to be saved in every check
seed: 3407                 # random seed number
patience: null            # patience for early stopping
unused_parameters: true   # needed for multi gpu case
best_model_criterion:     # criterion to save the best models
-   - train
    - total_count
    - max
cudnn_deterministic: false # setting to false accelerates the training speed but makes it non-deterministic
                           # in the case of GAN-TTS training, we strongly recommend setting to false
cudnn_benchmark: false     # setting to true might acdelerate the training speed but sometimes decrease it
                           # therefore, we set to false as a default (recommend trying both cases)