File size: 8,878 Bytes
12bfd03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
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
214
215
216
217
218
219
220
221
222
223
# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
import torch.nn as nn
from einops import pack, rearrange, repeat
from matcha.models.components.decoder import SinusoidalPosEmb, Block1D, ResnetBlock1D, Downsample1D, TimestepEmbedding, Upsample1D
from matcha.models.components.transformer import BasicTransformerBlock


class ConditionalDecoder(nn.Module):
    def __init__(
        self,
        in_channels,
        out_channels,
        channels=(256, 256),
        dropout=0.05,
        attention_head_dim=64,
        n_blocks=1,
        num_mid_blocks=2,
        num_heads=4,
        act_fn="snake",
    ):
        """
        This decoder requires an input with the same shape of the target. So, if your text content
        is shorter or longer than the outputs, please re-sampling it before feeding to the decoder.
        """
        super().__init__()
        channels = tuple(channels)
        self.in_channels = in_channels
        self.out_channels = out_channels

        self.time_embeddings = SinusoidalPosEmb(in_channels)
        time_embed_dim = channels[0] * 4
        self.time_mlp = TimestepEmbedding(
            in_channels=in_channels,
            time_embed_dim=time_embed_dim,
            act_fn="silu",
        )
        self.down_blocks = nn.ModuleList([])
        self.mid_blocks = nn.ModuleList([])
        self.up_blocks = nn.ModuleList([])

        output_channel = in_channels
        for i in range(len(channels)):  # pylint: disable=consider-using-enumerate
            input_channel = output_channel
            output_channel = channels[i]
            is_last = i == len(channels) - 1
            resnet = ResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
            transformer_blocks = nn.ModuleList(
                [
                    BasicTransformerBlock(
                        dim=output_channel,
                        num_attention_heads=num_heads,
                        attention_head_dim=attention_head_dim,
                        dropout=dropout,
                        activation_fn=act_fn,
                    )
                    for _ in range(n_blocks)
                ]
            )
            downsample = (
                Downsample1D(output_channel) if not is_last else nn.Conv1d(output_channel, output_channel, 3, padding=1)
            )
            self.down_blocks.append(nn.ModuleList([resnet, transformer_blocks, downsample]))

        for i in range(num_mid_blocks):
            input_channel = channels[-1]
            out_channels = channels[-1]
            resnet = ResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)

            transformer_blocks = nn.ModuleList(
                [
                    BasicTransformerBlock(
                        dim=output_channel,
                        num_attention_heads=num_heads,
                        attention_head_dim=attention_head_dim,
                        dropout=dropout,
                        activation_fn=act_fn,
                    )
                    for _ in range(n_blocks)
                ]
            )

            self.mid_blocks.append(nn.ModuleList([resnet, transformer_blocks]))

        channels = channels[::-1] + (channels[0],)
        for i in range(len(channels) - 1):
            input_channel = channels[i] * 2
            output_channel = channels[i + 1]
            is_last = i == len(channels) - 2
            resnet = ResnetBlock1D(
                dim=input_channel,
                dim_out=output_channel,
                time_emb_dim=time_embed_dim,
            )
            transformer_blocks = nn.ModuleList(
                [
                    BasicTransformerBlock(
                        dim=output_channel,
                        num_attention_heads=num_heads,
                        attention_head_dim=attention_head_dim,
                        dropout=dropout,
                        activation_fn=act_fn,
                    )
                    for _ in range(n_blocks)
                ]
            )
            upsample = (
                Upsample1D(output_channel, use_conv_transpose=True)
                if not is_last
                else nn.Conv1d(output_channel, output_channel, 3, padding=1)
            )
            self.up_blocks.append(nn.ModuleList([resnet, transformer_blocks, upsample]))
        self.final_block = Block1D(channels[-1], channels[-1])
        self.final_proj = nn.Conv1d(channels[-1], self.out_channels, 1)
        self.initialize_weights()


    def initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv1d):
                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)
            elif isinstance(m, nn.GroupNorm):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)
            elif isinstance(m, nn.Linear):
                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)

    def forward(self, x, mask, mu, t, spks=None, cond=None):
        """Forward pass of the UNet1DConditional model.

        Args:
            x (torch.Tensor): shape (batch_size, in_channels, time)
            mask (_type_): shape (batch_size, 1, time)
            t (_type_): shape (batch_size)
            spks (_type_, optional): shape: (batch_size, condition_channels). Defaults to None.
            cond (_type_, optional): placeholder for future use. Defaults to None.

        Raises:
            ValueError: _description_
            ValueError: _description_

        Returns:
            _type_: _description_
        """

        t = self.time_embeddings(t)
        t = self.time_mlp(t)

        x = pack([x, mu], "b * t")[0]

        if spks is not None:
            spks = repeat(spks, "b c -> b c t", t=x.shape[-1])
            x = pack([x, spks], "b * t")[0]
        if cond is not None:
            x = pack([x, cond], "b * t")[0]

        hiddens = []
        masks = [mask]
        for resnet, transformer_blocks, downsample in self.down_blocks:
            mask_down = masks[-1]
            x = resnet(x, mask_down, t)
            x = rearrange(x, "b c t -> b t c").contiguous()
            attn_mask = torch.matmul(mask_down.transpose(1, 2).contiguous(), mask_down)
            for transformer_block in transformer_blocks:
                x = transformer_block(
                    hidden_states=x,
                    attention_mask=attn_mask,
                    timestep=t,
                )
            x = rearrange(x, "b t c -> b c t").contiguous()
            hiddens.append(x)  # Save hidden states for skip connections
            x = downsample(x * mask_down)
            masks.append(mask_down[:, :, ::2])
        masks = masks[:-1]
        mask_mid = masks[-1]

        for resnet, transformer_blocks in self.mid_blocks:
            x = resnet(x, mask_mid, t)
            x = rearrange(x, "b c t -> b t c").contiguous()
            attn_mask = torch.matmul(mask_mid.transpose(1, 2).contiguous(), mask_mid)
            for transformer_block in transformer_blocks:
                x = transformer_block(
                    hidden_states=x,
                    attention_mask=attn_mask,
                    timestep=t,
                )
            x = rearrange(x, "b t c -> b c t").contiguous()

        for resnet, transformer_blocks, upsample in self.up_blocks:
            mask_up = masks.pop()
            skip = hiddens.pop()
            x = pack([x[:, :, :skip.shape[-1]], skip], "b * t")[0]
            x = resnet(x, mask_up, t)
            x = rearrange(x, "b c t -> b t c").contiguous()
            attn_mask = torch.matmul(mask_up.transpose(1, 2).contiguous(), mask_up)
            for transformer_block in transformer_blocks:
                x = transformer_block(
                    hidden_states=x,
                    attention_mask=attn_mask,
                    timestep=t,
                )
            x = rearrange(x, "b t c -> b c t").contiguous()
            x = upsample(x * mask_up)
        x = self.final_block(x, mask_up)
        output = self.final_proj(x * mask_up)
        return output * mask