File size: 13,666 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
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
"""Encodec SEANet-based encoder and decoder implementation."""
import typing as tp

import numpy as np
import torch.nn as nn

from academicodec.modules import SConv1d
from academicodec.modules import SConvTranspose1d
from academicodec.modules import SLSTM


class SEANetResnetBlock(nn.Module):
    """Residual block from SEANet model.
    Args:
        dim (int): Dimension of the input/output
        kernel_sizes (list): List of kernel sizes for the convolutions.
        dilations (list): List of dilations for the convolutions.
        activation (str): Activation function.
        activation_params (dict): Parameters to provide to the activation function
        norm (str): Normalization method.
        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
        causal (bool): Whether to use fully causal convolution.
        pad_mode (str): Padding mode for the convolutions.
        compress (int): Reduced dimensionality in residual branches (from Demucs v3)
        true_skip (bool): Whether to use true skip connection or a simple convolution as the skip connection.
    """

    def __init__(self,
                 dim: int,
                 kernel_sizes: tp.List[int]=[3, 1],
                 dilations: tp.List[int]=[1, 1],
                 activation: str='ELU',
                 activation_params: dict={'alpha': 1.0},
                 norm: str='weight_norm',
                 norm_params: tp.Dict[str, tp.Any]={},
                 causal: bool=False,
                 pad_mode: str='reflect',
                 compress: int=2,
                 true_skip: bool=True):
        super().__init__()
        assert len(kernel_sizes) == len(
            dilations), 'Number of kernel sizes should match number of dilations'
        act = getattr(nn, activation)
        hidden = dim // compress
        block = []
        for i, (kernel_size,
                dilation) in enumerate(zip(kernel_sizes, dilations)):
            in_chs = dim if i == 0 else hidden
            out_chs = dim if i == len(kernel_sizes) - 1 else hidden
            block += [
                act(**activation_params),
                SConv1d(
                    in_chs,
                    out_chs,
                    kernel_size=kernel_size,
                    dilation=dilation,
                    norm=norm,
                    norm_kwargs=norm_params,
                    causal=causal,
                    pad_mode=pad_mode),
            ]
        self.block = nn.Sequential(*block)
        self.shortcut: nn.Module
        if true_skip:
            self.shortcut = nn.Identity()
        else:
            self.shortcut = SConv1d(
                dim,
                dim,
                kernel_size=1,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode)

    def forward(self, x):
        return self.shortcut(x) + self.block(x)


class SEANetEncoder(nn.Module):
    """SEANet encoder.
    Args:
        channels (int): Audio channels.
        dimension (int): Intermediate representation dimension.
        n_filters (int): Base width for the model.
        n_residual_layers (int): nb of residual layers.
        ratios (Sequence[int]): kernel size and stride ratios. The encoder uses downsampling ratios instead of
            upsampling ratios, hence it will use the ratios in the reverse order to the ones specified here
            that must match the decoder order
        activation (str): Activation function.
        activation_params (dict): Parameters to provide to the activation function
        norm (str): Normalization method.
        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
        kernel_size (int): Kernel size for the initial convolution.
        last_kernel_size (int): Kernel size for the initial convolution.
        residual_kernel_size (int): Kernel size for the residual layers.
        dilation_base (int): How much to increase the dilation with each layer.
        causal (bool): Whether to use fully causal convolution.
        pad_mode (str): Padding mode for the convolutions.
        true_skip (bool): Whether to use true skip connection or a simple
            (streamable) convolution as the skip connection in the residual network blocks.
        compress (int): Reduced dimensionality in residual branches (from Demucs v3).
        lstm (int): Number of LSTM layers at the end of the encoder.
    """

    def __init__(self,
                 channels: int=1,
                 dimension: int=128,
                 n_filters: int=32,
                 n_residual_layers: int=1,
                 ratios: tp.List[int]=[8, 5, 4, 2],
                 activation: str='ELU',
                 activation_params: dict={'alpha': 1.0},
                 norm: str='weight_norm',
                 norm_params: tp.Dict[str, tp.Any]={},
                 kernel_size: int=7,
                 last_kernel_size: int=7,
                 residual_kernel_size: int=3,
                 dilation_base: int=2,
                 causal: bool=False,
                 pad_mode: str='reflect',
                 true_skip: bool=False,
                 compress: int=2,
                 lstm: int=2):
        super().__init__()
        self.channels = channels
        self.dimension = dimension
        self.n_filters = n_filters
        self.ratios = list(reversed(ratios))
        del ratios
        self.n_residual_layers = n_residual_layers
        self.hop_length = np.prod(self.ratios)  # 计算乘积

        act = getattr(nn, activation)
        mult = 1
        model: tp.List[nn.Module] = [
            SConv1d(
                channels,
                mult * n_filters,
                kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode)
        ]
        # Downsample to raw audio scale
        for i, ratio in enumerate(self.ratios):
            # Add residual layers
            for j in range(n_residual_layers):
                model += [
                    SEANetResnetBlock(
                        mult * n_filters,
                        kernel_sizes=[residual_kernel_size, 1],
                        dilations=[dilation_base**j, 1],
                        norm=norm,
                        norm_params=norm_params,
                        activation=activation,
                        activation_params=activation_params,
                        causal=causal,
                        pad_mode=pad_mode,
                        compress=compress,
                        true_skip=true_skip)
                ]

            # Add downsampling layers
            model += [
                act(**activation_params),
                SConv1d(
                    mult * n_filters,
                    mult * n_filters * 2,
                    kernel_size=ratio * 2,
                    stride=ratio,
                    norm=norm,
                    norm_kwargs=norm_params,
                    causal=causal,
                    pad_mode=pad_mode),
            ]
            mult *= 2

        if lstm:
            model += [SLSTM(mult * n_filters, num_layers=lstm)]

        model += [
            act(**activation_params), SConv1d(
                mult * n_filters,
                dimension,
                last_kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode)
        ]

        self.model = nn.Sequential(*model)

    def forward(self, x):
        return self.model(x)


class SEANetDecoder(nn.Module):
    """SEANet decoder.
    Args:
        channels (int): Audio channels.
        dimension (int): Intermediate representation dimension.
        n_filters (int): Base width for the model.
        n_residual_layers (int): nb of residual layers.
        ratios (Sequence[int]): kernel size and stride ratios
        activation (str): Activation function.
        activation_params (dict): Parameters to provide to the activation function
        final_activation (str): Final activation function after all convolutions.
        final_activation_params (dict): Parameters to provide to the activation function
        norm (str): Normalization method.
        norm_params (dict): Parameters to provide to the underlying normalization used along with the convolution.
        kernel_size (int): Kernel size for the initial convolution.
        last_kernel_size (int): Kernel size for the initial convolution.
        residual_kernel_size (int): Kernel size for the residual layers.
        dilation_base (int): How much to increase the dilation with each layer.
        causal (bool): Whether to use fully causal convolution.
        pad_mode (str): Padding mode for the convolutions.
        true_skip (bool): Whether to use true skip connection or a simple
            (streamable) convolution as the skip connection in the residual network blocks.
        compress (int): Reduced dimensionality in residual branches (from Demucs v3).
        lstm (int): Number of LSTM layers at the end of the encoder.
        trim_right_ratio (float): Ratio for trimming at the right of the transposed convolution under the causal setup.
            If equal to 1.0, it means that all the trimming is done at the right.
    """

    def __init__(self,
                 channels: int=1,
                 dimension: int=128,
                 n_filters: int=32,
                 n_residual_layers: int=1,
                 ratios: tp.List[int]=[8, 5, 4, 2],
                 activation: str='ELU',
                 activation_params: dict={'alpha': 1.0},
                 final_activation: tp.Optional[str]=None,
                 final_activation_params: tp.Optional[dict]=None,
                 norm: str='weight_norm',
                 norm_params: tp.Dict[str, tp.Any]={},
                 kernel_size: int=7,
                 last_kernel_size: int=7,
                 residual_kernel_size: int=3,
                 dilation_base: int=2,
                 causal: bool=False,
                 pad_mode: str='reflect',
                 true_skip: bool=False,
                 compress: int=2,
                 lstm: int=2,
                 trim_right_ratio: float=1.0):
        super().__init__()
        self.dimension = dimension
        self.channels = channels
        self.n_filters = n_filters
        self.ratios = ratios
        del ratios
        self.n_residual_layers = n_residual_layers
        self.hop_length = np.prod(self.ratios)

        act = getattr(nn, activation)
        mult = int(2**len(self.ratios))
        model: tp.List[nn.Module] = [
            SConv1d(
                dimension,
                mult * n_filters,
                kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode)
        ]

        if lstm:
            model += [SLSTM(mult * n_filters, num_layers=lstm)]

        # Upsample to raw audio scale
        for i, ratio in enumerate(self.ratios):
            # Add upsampling layers
            model += [
                act(**activation_params),
                SConvTranspose1d(
                    mult * n_filters,
                    mult * n_filters // 2,
                    kernel_size=ratio * 2,
                    stride=ratio,
                    norm=norm,
                    norm_kwargs=norm_params,
                    causal=causal,
                    trim_right_ratio=trim_right_ratio),
            ]
            # Add residual layers
            for j in range(n_residual_layers):
                model += [
                    SEANetResnetBlock(
                        mult * n_filters // 2,
                        kernel_sizes=[residual_kernel_size, 1],
                        dilations=[dilation_base**j, 1],
                        activation=activation,
                        activation_params=activation_params,
                        norm=norm,
                        norm_params=norm_params,
                        causal=causal,
                        pad_mode=pad_mode,
                        compress=compress,
                        true_skip=true_skip)
                ]

            mult //= 2

        # Add final layers
        model += [
            act(**activation_params), SConv1d(
                n_filters,
                channels,
                last_kernel_size,
                norm=norm,
                norm_kwargs=norm_params,
                causal=causal,
                pad_mode=pad_mode)
        ]
        # Add optional final activation to decoder (eg. tanh)
        if final_activation is not None:
            final_act = getattr(nn, final_activation)
            final_activation_params = final_activation_params or {}
            model += [final_act(**final_activation_params)]
        self.model = nn.Sequential(*model)

    def forward(self, z):
        y = self.model(z)
        return y


def test():
    import torch
    encoder = SEANetEncoder()
    decoder = SEANetDecoder()
    x = torch.randn(1, 1, 24000)
    z = encoder(x)
    print('z ', z.shape)
    assert 1 == 2
    assert list(z.shape) == [1, 128, 75], z.shape
    y = decoder(z)
    assert y.shape == x.shape, (x.shape, y.shape)


if __name__ == '__main__':
    test()