SafeSpeech/toolbox.py at master · wxzyd123/SafeSpeech · 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
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
import torch

from bert_vits2.text.symbols import symbols


def build_models_noise(hps, device):
    '''
        Build models for perturbation genetation.
    '''
    from bert_vits2.models_noise import (
        SynthesizerTrn,
        MultiPeriodDiscriminator,
        DurationDiscriminator,
        WavLMDiscriminator,
    )

    mas_noise_scale_initial = 0.01
    noise_scale_delta = 2e-6

    net_dur_disc = DurationDiscriminator(
        hps.model.hidden_channels,
        hps.model.hidden_channels,
        3,
        0.1,
        gin_channels=hps.model.gin_channels if hps.data.n_speakers != 0 else 0,
    ).to(device)

    net_g = SynthesizerTrn(
        len(symbols),
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        n_speakers=hps.data.n_speakers,
        mas_noise_scale_initial=mas_noise_scale_initial,
        noise_scale_delta=noise_scale_delta,
        **hps.model,
    ).to(device)

    for param in net_g.enc_p.bert_proj.parameters():
        param.requires_grad = False
    for param in net_g.enc_p.ja_bert_proj.parameters():
        param.requires_grad = False

    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).to(device)
    net_wd = WavLMDiscriminator(
        hps.model.slm.hidden, hps.model.slm.nlayers, hps.model.slm.initial_channel
    ).to(device)

    return net_g, net_d, net_wd, net_dur_disc


def build_models(hps, device):
    '''
        Builde models for fine-tuning [This is the original model without modification.]
    '''

    from bert_vits2.models import (
        SynthesizerTrn,
        MultiPeriodDiscriminator,
        DurationDiscriminator,
        WavLMDiscriminator,
    )

    mas_noise_scale_initial = 0.01
    noise_scale_delta = 2e-6

    net_dur_disc = DurationDiscriminator(
        hps.model.hidden_channels,
        hps.model.hidden_channels,
        3,
        0.1,
        gin_channels=hps.model.gin_channels if hps.data.n_speakers != 0 else 0,
    ).to(device)

    net_g = SynthesizerTrn(
        len(symbols),
        hps.data.filter_length // 2 + 1,
        hps.train.segment_size // hps.data.hop_length,
        n_speakers=hps.data.n_speakers,
        mas_noise_scale_initial=mas_noise_scale_initial,
        noise_scale_delta=noise_scale_delta,
        **hps.model,
    ).to(device)

    for param in net_g.enc_p.bert_proj.parameters():
        param.requires_grad = False
    for param in net_g.enc_p.ja_bert_proj.parameters():
        param.requires_grad = False

    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).to(device)
    net_wd = WavLMDiscriminator(
        hps.model.slm.hidden, hps.model.slm.nlayers, hps.model.slm.initial_channel
    ).to(device)

    return net_g, net_d, net_wd, net_dur_disc


def build_optims(hps, nets):
    '''
        Build the optimizers for fine-tuning
    '''

    net_g, net_d, net_wd, net_dur_disc = nets
    optim_g = torch.optim.AdamW(
        filter(lambda p: p.requires_grad, net_g.parameters()),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )
    optim_d = torch.optim.AdamW(
        net_d.parameters(),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )
    optim_wd = torch.optim.AdamW(
        net_wd.parameters(),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )
    optim_dur_disc = torch.optim.AdamW(
        net_dur_disc.parameters(),
        hps.train.learning_rate,
        betas=hps.train.betas,
        eps=hps.train.eps,
    )

    return optim_g, optim_d, optim_wd, optim_dur_disc


def build_schedulers(hps, optims, epoch_str):
    '''
        Build the schedulers for optimizers when fine-tuning.
    '''
    optim_g, optim_d, optim_wd, optim_dur_disc = optims
    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
        optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    )
    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
        optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    )
    scheduler_wd = torch.optim.lr_scheduler.ExponentialLR(
        optim_wd, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    )
    scheduler_dur_disc = torch.optim.lr_scheduler.ExponentialLR(
        optim_dur_disc, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
    )

    return scheduler_g, scheduler_d, scheduler_wd, scheduler_dur_disc