unet_grad_tts.py 6.04 KB
Newer Older
patil-suraj's avatar
patil-suraj committed
1
2
3
4
import torch

from ..configuration_utils import ConfigMixin
from ..modeling_utils import ModelMixin
Patrick von Platen's avatar
Patrick von Platen committed
5
from .attention import LinearAttention
6
from .embeddings import get_timestep_embedding
patil-suraj's avatar
patil-suraj committed
7
from .resnet import Downsample, Upsample
patil-suraj's avatar
patil-suraj committed
8

9

patil-suraj's avatar
patil-suraj committed
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
class Mish(torch.nn.Module):
    def forward(self, x):
        return x * torch.tanh(torch.nn.functional.softplus(x))


class Rezero(torch.nn.Module):
    def __init__(self, fn):
        super(Rezero, self).__init__()
        self.fn = fn
        self.g = torch.nn.Parameter(torch.zeros(1))

    def forward(self, x):
        return self.fn(x) * self.g


class Block(torch.nn.Module):
    def __init__(self, dim, dim_out, groups=8):
        super(Block, self).__init__()
28
29
30
        self.block = torch.nn.Sequential(
            torch.nn.Conv2d(dim, dim_out, 3, padding=1), torch.nn.GroupNorm(groups, dim_out), Mish()
        )
patil-suraj's avatar
patil-suraj committed
31
32
33
34
35
36
37
38
39

    def forward(self, x, mask):
        output = self.block(x * mask)
        return output * mask


class ResnetBlock(torch.nn.Module):
    def __init__(self, dim, dim_out, time_emb_dim, groups=8):
        super(ResnetBlock, self).__init__()
40
        self.mlp = torch.nn.Sequential(Mish(), torch.nn.Linear(time_emb_dim, dim_out))
patil-suraj's avatar
patil-suraj committed
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

        self.block1 = Block(dim, dim_out, groups=groups)
        self.block2 = Block(dim_out, dim_out, groups=groups)
        if dim != dim_out:
            self.res_conv = torch.nn.Conv2d(dim, dim_out, 1)
        else:
            self.res_conv = torch.nn.Identity()

    def forward(self, x, mask, time_emb):
        h = self.block1(x, mask)
        h += self.mlp(time_emb).unsqueeze(-1).unsqueeze(-1)
        h = self.block2(h, mask)
        output = h + self.res_conv(x * mask)
        return output


class Residual(torch.nn.Module):
    def __init__(self, fn):
        super(Residual, self).__init__()
        self.fn = fn

    def forward(self, x, *args, **kwargs):
        output = self.fn(x, *args, **kwargs) + x
        return output


class UNetGradTTSModel(ModelMixin, ConfigMixin):
68
    def __init__(self, dim, dim_mults=(1, 2, 4), groups=8, n_spks=None, spk_emb_dim=64, n_feats=80, pe_scale=1000):
patil-suraj's avatar
patil-suraj committed
69
70
        super(UNetGradTTSModel, self).__init__()

71
        self.register_to_config(
patil-suraj's avatar
patil-suraj committed
72
73
74
75
76
77
            dim=dim,
            dim_mults=dim_mults,
            groups=groups,
            n_spks=n_spks,
            spk_emb_dim=spk_emb_dim,
            n_feats=n_feats,
78
            pe_scale=pe_scale,
patil-suraj's avatar
patil-suraj committed
79
        )
80

patil-suraj's avatar
patil-suraj committed
81
82
83
84
85
86
        self.dim = dim
        self.dim_mults = dim_mults
        self.groups = groups
        self.n_spks = n_spks if not isinstance(n_spks, type(None)) else 1
        self.spk_emb_dim = spk_emb_dim
        self.pe_scale = pe_scale
87

patil-suraj's avatar
patil-suraj committed
88
        if n_spks > 1:
patil-suraj's avatar
patil-suraj committed
89
            self.spk_emb = torch.nn.Embedding(n_spks, spk_emb_dim)
patil-suraj's avatar
style  
patil-suraj committed
90
91
92
            self.spk_mlp = torch.nn.Sequential(
                torch.nn.Linear(spk_emb_dim, spk_emb_dim * 4), Mish(), torch.nn.Linear(spk_emb_dim * 4, n_feats)
            )
93

94
        self.mlp = torch.nn.Sequential(torch.nn.Linear(dim, dim * 4), Mish(), torch.nn.Linear(dim * 4, dim))
patil-suraj's avatar
patil-suraj committed
95
96
97
98
99
100
101
102
103

        dims = [2 + (1 if n_spks > 1 else 0), *map(lambda m: dim * m, dim_mults)]
        in_out = list(zip(dims[:-1], dims[1:]))
        self.downs = torch.nn.ModuleList([])
        self.ups = torch.nn.ModuleList([])
        num_resolutions = len(in_out)

        for ind, (dim_in, dim_out) in enumerate(in_out):
            is_last = ind >= (num_resolutions - 1)
104
105
106
107
108
109
            self.downs.append(
                torch.nn.ModuleList(
                    [
                        ResnetBlock(dim_in, dim_out, time_emb_dim=dim),
                        ResnetBlock(dim_out, dim_out, time_emb_dim=dim),
                        Residual(Rezero(LinearAttention(dim_out))),
patil-suraj's avatar
patil-suraj committed
110
                        Downsample(dim_out, use_conv=True, padding=1) if not is_last else torch.nn.Identity(),
111
112
113
                    ]
                )
            )
patil-suraj's avatar
patil-suraj committed
114
115
116
117
118
119
120

        mid_dim = dims[-1]
        self.mid_block1 = ResnetBlock(mid_dim, mid_dim, time_emb_dim=dim)
        self.mid_attn = Residual(Rezero(LinearAttention(mid_dim)))
        self.mid_block2 = ResnetBlock(mid_dim, mid_dim, time_emb_dim=dim)

        for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
121
122
123
124
125
126
            self.ups.append(
                torch.nn.ModuleList(
                    [
                        ResnetBlock(dim_out * 2, dim_in, time_emb_dim=dim),
                        ResnetBlock(dim_in, dim_in, time_emb_dim=dim),
                        Residual(Rezero(LinearAttention(dim_in))),
patil-suraj's avatar
patil-suraj committed
127
                        Upsample(dim_in, use_conv_transpose=True),
128
129
130
                    ]
                )
            )
patil-suraj's avatar
patil-suraj committed
131
132
133
        self.final_block = Block(dim, dim)
        self.final_conv = torch.nn.Conv2d(dim, 1, 1)

patil-suraj's avatar
patil-suraj committed
134
    def forward(self, x, timesteps, mu, mask, spk=None):
patil-suraj's avatar
patil-suraj committed
135
136
137
138
        if self.n_spks > 1:
            # Get speaker embedding
            spk = self.spk_emb(spk)

patil-suraj's avatar
patil-suraj committed
139
140
        if not isinstance(spk, type(None)):
            s = self.spk_mlp(spk)
141

142
        t = get_timestep_embedding(timesteps, self.dim, scale=self.pe_scale)
patil-suraj's avatar
patil-suraj committed
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
        t = self.mlp(t)

        if self.n_spks < 2:
            x = torch.stack([mu, x], 1)
        else:
            s = s.unsqueeze(-1).repeat(1, 1, x.shape[-1])
            x = torch.stack([mu, x, s], 1)
        mask = mask.unsqueeze(1)

        hiddens = []
        masks = [mask]
        for resnet1, resnet2, attn, downsample in self.downs:
            mask_down = masks[-1]
            x = resnet1(x, mask_down, t)
            x = resnet2(x, mask_down, t)
            x = attn(x)
            hiddens.append(x)
            x = downsample(x * mask_down)
            masks.append(mask_down[:, :, :, ::2])

        masks = masks[:-1]
        mask_mid = masks[-1]
        x = self.mid_block1(x, mask_mid, t)
        x = self.mid_attn(x)
        x = self.mid_block2(x, mask_mid, t)

        for resnet1, resnet2, attn, upsample in self.ups:
            mask_up = masks.pop()
            x = torch.cat((x, hiddens.pop()), dim=1)
            x = resnet1(x, mask_up, t)
            x = resnet2(x, mask_up, t)
            x = attn(x)
            x = upsample(x * mask_up)

        x = self.final_block(x, mask)
        output = self.final_conv(x * mask)

180
        return (output * mask).squeeze(1)