unet_grad_tts.py 7.05 KB
Newer Older
patil-suraj's avatar
patil-suraj committed
1
2
import torch

3

patil-suraj's avatar
patil-suraj committed
4
try:
patil-suraj's avatar
patil-suraj committed
5
    from einops import rearrange
patil-suraj's avatar
patil-suraj committed
6
7
8
9
10
11
except:
    print("Einops is not installed")
    pass

from ..configuration_utils import ConfigMixin
from ..modeling_utils import ModelMixin
12
from .embeddings import get_timestep_embedding
patil-suraj's avatar
patil-suraj committed
13
from .resnet import Upsample
patil-suraj's avatar
patil-suraj committed
14

15

patil-suraj's avatar
patil-suraj committed
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
class Mish(torch.nn.Module):
    def forward(self, x):
        return x * torch.tanh(torch.nn.functional.softplus(x))


class Downsample(torch.nn.Module):
    def __init__(self, dim):
        super(Downsample, self).__init__()
        self.conv = torch.nn.Conv2d(dim, dim, 3, 2, 1)

    def forward(self, x):
        return self.conv(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__()
43
44
45
        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
46
47
48
49
50
51
52
53
54

    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__()
55
        self.mlp = torch.nn.Sequential(Mish(), torch.nn.Linear(time_emb_dim, dim_out))
patil-suraj's avatar
patil-suraj committed
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77

        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 LinearAttention(torch.nn.Module):
    def __init__(self, dim, heads=4, dim_head=32):
        super(LinearAttention, self).__init__()
        self.heads = heads
        hidden_dim = dim_head * heads
        self.to_qkv = torch.nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
78
        self.to_out = torch.nn.Conv2d(hidden_dim, dim, 1)
patil-suraj's avatar
patil-suraj committed
79
80
81
82

    def forward(self, x):
        b, c, h, w = x.shape
        qkv = self.to_qkv(x)
83
        q, k, v = rearrange(qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3)
patil-suraj's avatar
patil-suraj committed
84
        k = k.softmax(dim=-1)
85
86
87
        context = torch.einsum("bhdn,bhen->bhde", k, v)
        out = torch.einsum("bhde,bhdn->bhen", context, q)
        out = rearrange(out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w)
patil-suraj's avatar
patil-suraj committed
88
89
90
91
92
93
94
95
96
97
98
99
100
101
        return self.to_out(out)


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):
102
    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
103
104
        super(UNetGradTTSModel, self).__init__()

105
        self.register_to_config(
patil-suraj's avatar
patil-suraj committed
106
107
108
109
110
111
            dim=dim,
            dim_mults=dim_mults,
            groups=groups,
            n_spks=n_spks,
            spk_emb_dim=spk_emb_dim,
            n_feats=n_feats,
112
            pe_scale=pe_scale,
patil-suraj's avatar
patil-suraj committed
113
        )
114

patil-suraj's avatar
patil-suraj committed
115
116
117
118
119
120
        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
121

patil-suraj's avatar
patil-suraj committed
122
        if n_spks > 1:
patil-suraj's avatar
patil-suraj committed
123
            self.spk_emb = torch.nn.Embedding(n_spks, spk_emb_dim)
patil-suraj's avatar
style  
patil-suraj committed
124
125
126
            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)
            )
127

128
        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
129
130
131
132
133
134
135
136
137

        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)
138
139
140
141
142
143
144
145
146
147
            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))),
                        Downsample(dim_out) if not is_last else torch.nn.Identity(),
                    ]
                )
            )
patil-suraj's avatar
patil-suraj committed
148
149
150
151
152
153
154

        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:])):
155
156
157
158
159
160
            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
161
                        Upsample(dim_in, use_conv_transpose=True),
162
163
164
                    ]
                )
            )
patil-suraj's avatar
patil-suraj committed
165
166
167
        self.final_block = Block(dim, dim)
        self.final_conv = torch.nn.Conv2d(dim, 1, 1)

patil-suraj's avatar
patil-suraj committed
168
    def forward(self, x, timesteps, mu, mask, spk=None):
patil-suraj's avatar
patil-suraj committed
169
170
171
172
        if self.n_spks > 1:
            # Get speaker embedding
            spk = self.spk_emb(spk)

patil-suraj's avatar
patil-suraj committed
173
174
        if not isinstance(spk, type(None)):
            s = self.spk_mlp(spk)
175

176
        t = get_timestep_embedding(timesteps, self.dim, scale=self.pe_scale)
patil-suraj's avatar
patil-suraj committed
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
        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)

214
        return (output * mask).squeeze(1)