File size: 16,159 Bytes
df9f13e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
import math
import time
from collections import OrderedDict
from typing import Dict, List, Optional, Tuple
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from espnet2.torch_utils.get_layer_from_string import get_layer
from torch.nn import init
from torch.nn.parameter import Parameter
import src.utils as utils


class Lambda(nn.Module):
    def __init__(self, lambd):
        super().__init__()
        import types

        assert type(lambd) is types.LambdaType
        self.lambd = lambd

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


class LayerNormPermuted(nn.LayerNorm):
    def __init__(self, *args, **kwargs):
        super(LayerNormPermuted, self).__init__(*args, **kwargs)

    def forward(self, x):
        """
        Args:
            x: [B, C, T, F]
        """
        x = x.permute(0, 2, 3, 1)  # [B, T, F, C]
        x = super().forward(x)
        x = x.permute(0, 3, 1, 2)  # [B, C, T, F]
        return x


# Use native layernorm implementation
class LayerNormalization4D(nn.Module):
    def __init__(self, C, eps=1e-5, preserve_outdim=False):
        super().__init__()
        self.norm = nn.LayerNorm(C, eps=eps)
        self.preserve_outdim = preserve_outdim

    def forward(self, x: torch.Tensor):
        """
        input: (*, C)
        """
        x = self.norm(x)
        return x


class LayerNormalization4DCF(nn.Module):
    def __init__(self, input_dimension, eps=1e-5):
        assert len(input_dimension) == 2
        Q, C = input_dimension
        super().__init__()
        self.norm = nn.LayerNorm((Q * C), eps=eps)

    def forward(self, x: torch.Tensor):
        """
        input: (B, T, Q * C)
        """
        x = self.norm(x)

        return x


class LayerNormalization4D_old(nn.Module):
    def __init__(self, input_dimension, eps=1e-5):
        super().__init__()
        param_size = [1, input_dimension, 1, 1]
        self.gamma = Parameter(torch.Tensor(*param_size).to(torch.float32))
        self.beta = Parameter(torch.Tensor(*param_size).to(torch.float32))
        init.ones_(self.gamma)
        init.zeros_(self.beta)
        self.eps = eps

    def forward(self, x):
        if x.ndim == 4:
            _, C, _, _ = x.shape
            stat_dim = (1,)
        else:
            raise ValueError("Expect x to have 4 dimensions, but got {}".format(x.ndim))
        mu_ = x.mean(dim=stat_dim, keepdim=True)  # [B,1,T,F]
        std_ = torch.sqrt(x.var(dim=stat_dim, unbiased=False, keepdim=True) + self.eps)  # [B,1,T,F]
        x_hat = ((x - mu_) / std_) * self.gamma + self.beta
        return x_hat


def mod_pad(x, chunk_size, pad):
    # Mod pad the rminput to perform integer number of
    # inferences
    mod = 0
    if (x.shape[-1] % chunk_size) != 0:
        mod = chunk_size - (x.shape[-1] % chunk_size)

    x = F.pad(x, (0, mod))
    x = F.pad(x, pad)

    return x, mod


class Attention_STFT_causal(nn.Module):
    def __getitem__(self, key):
        return getattr(self, key)

    def __init__(
        self,
        emb_dim,
        n_freqs,
        approx_qk_dim=512,
        n_head=4,
        activation="prelu",
        eps=1e-5,
        skip_conn=True,
        use_flash_attention=False,
        dim_feedforward=-1,
    ):
        super().__init__()
        self.position_code = utils.PositionalEncoding(emb_dim * n_freqs, max_len=5000)

        self.skip_conn = skip_conn
        self.n_freqs = n_freqs
        self.E = math.ceil(approx_qk_dim * 1.0 / n_freqs)  # approx_qk_dim is only approximate
        self.n_head = n_head
        self.V_dim = emb_dim // n_head
        self.emb_dim = emb_dim
        assert emb_dim % n_head == 0
        E = self.E

        self.add_module(
            "attn_conv_Q",
            nn.Sequential(
                nn.Linear(emb_dim, E * n_head),  # [B, T, Q, HE]
                get_layer(activation)(),
                # [B, T, Q, H, E] -> [B, H, T, Q, E] ->  [B * H, T, Q * E]
                Lambda(
                    lambda x: x.reshape(x.shape[0], x.shape[1], x.shape[2], n_head, E)
                    .permute(0, 3, 1, 2, 4)
                    .reshape(x.shape[0] * n_head, x.shape[1], x.shape[2] * E)
                ),  # (BH, T, Q * E)
                LayerNormalization4DCF((n_freqs, E), eps=eps),
            ),
        )
        self.add_module(
            "attn_conv_K",
            nn.Sequential(
                nn.Linear(emb_dim, E * n_head),
                get_layer(activation)(),
                Lambda(
                    lambda x: x.reshape(x.shape[0], x.shape[1], x.shape[2], n_head, E)
                    .permute(0, 3, 1, 2, 4)
                    .reshape(x.shape[0] * n_head, x.shape[1], x.shape[2] * E)
                ),
                LayerNormalization4DCF((n_freqs, E), eps=eps),
            ),
        )
        self.add_module(
            "attn_conv_V",
            nn.Sequential(
                nn.Linear(emb_dim, (emb_dim // n_head) * n_head),
                get_layer(activation)(),
                Lambda(
                    lambda x: x.reshape(x.shape[0], x.shape[1], x.shape[2], n_head, (emb_dim // n_head))
                    .permute(0, 3, 1, 2, 4)
                    .reshape(x.shape[0] * n_head, x.shape[1], x.shape[2] * (emb_dim // n_head))
                ),
                LayerNormalization4DCF((n_freqs, emb_dim // n_head), eps=eps),
            ),
        )

        self.dim_feedforward = dim_feedforward

        if dim_feedforward == -1:
            self.add_module(
                "attn_concat_proj",
                nn.Sequential(
                    nn.Linear(emb_dim, emb_dim),
                    get_layer(activation)(),
                    Lambda(lambda x: x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3])),
                    LayerNormalization4DCF((n_freqs, emb_dim), eps=eps),
                ),
            )
        else:
            self.linear1 = nn.Linear(emb_dim, dim_feedforward)
            self.dropout = nn.Dropout(p=0.1)
            self.activation = nn.ReLU()
            self.linear2 = nn.Linear(dim_feedforward, emb_dim)
            self.dropout2 = nn.Dropout(p=0.1)
            self.norm = LayerNormalization4DCF((n_freqs, emb_dim), eps=eps)

    def _ff_block(self, x):
        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
        return self.dropout2(x)

    def get_lookahead_mask(self, seq_len, device):
        """Creates a binary mask for each sequence which masks future frames.
        Arguments
        ---------
        seq_len: int
            Length of the sequence.
        device: torch.device
            The device on which to create the mask.
        Example
        -------
        >>> a = torch.LongTensor([[1,1,0], [2,3,0], [4,5,0]])
        >>> get_lookahead_mask(a.shape[1], device)
        tensor([[0., -inf, -inf],
                [0., 0., -inf],
                [0., 0., 0.]])
        """
        mask = (torch.triu(torch.ones((seq_len, seq_len), device=device)) == 1).transpose(0, 1)

        return mask.detach().to(device)

    def forward(self, batch):
        ### input/output B T F C
        # attention
        inputs = batch
        B0, T0, Q0, C0 = batch.shape
        # print("dim of just entering attention stft causal is {}".format(batch.shape))
        # [2, 12, 133, 16]

        # positional encoding
        pos_code = self.position_code(batch)  # 1, T, embed_dim
        # print("pos_code", pos_code.shape)
        _, T, QC = pos_code.shape
        pos_code = pos_code.reshape(1, T, Q0, C0)
        batch = batch + pos_code

        # print("shape of q is {}".format(Q.shape))
        # print("batch shape is {}".format(batch.shape)) [1, 4800, 16, 133]

        Q = self["attn_conv_Q"](batch)  # [B', T, Q * C]
        K = self["attn_conv_K"](batch)  # [B', T, Q * C]
        V = self["attn_conv_V"](batch)  # [B', T, Q * C]

        emb_dim = Q.shape[-1]

        local_mask = self.get_lookahead_mask(batch.shape[1], batch.device)

        attn_mat = torch.matmul(Q, K.transpose(1, 2)) / (emb_dim**0.5)  # [B', T, T]
        attn_mat.masked_fill_(local_mask == 0, -float("Inf"))
        attn_mat = F.softmax(attn_mat, dim=2)  # [B', T, T]

        V = torch.matmul(attn_mat, V)  # [B', T, Q*C]
        V = V.reshape(-1, T0, V.shape[-1])  # [BH, T, Q * C]
        V = V.transpose(1, 2)  # [B', Q * C, T]

        batch = V.reshape(B0, self.n_head, self.n_freqs, self.V_dim, T0)  # [B, H, Q, C, T]
        batch = batch.transpose(2, 3)  # [B, H, C, Q, T]
        batch = batch.reshape(B0, self.n_head * self.V_dim, self.n_freqs, T0)  # [B, HC, Q, T]
        batch = batch.permute(0, 3, 2, 1)  # [B, T, Q, C]

        if self.dim_feedforward == -1:
            batch = self["attn_concat_proj"](batch)  # [B, T, Q * C]
        else:
            batch = batch + self._ff_block(batch)  # [B, T, Q, C]
            batch = batch.reshape(batch.shape[0], batch.shape[1], batch.shape[2] * batch.shape[3])
            batch = self.norm(batch)
        batch = batch.reshape(batch.shape[0], batch.shape[1], Q0, C0)  # [B, T, Q, C])

        # print("dim of output of attention stft causal is {}".format(batch.shape))
        # [2, 12, 133, 16]

        # Add batch if attention is performed
        if self.skip_conn:
            return batch + inputs
        else:
            return batch


class GridNetBlock(nn.Module):
    def __getitem__(self, key):
        return getattr(self, key)

    def __init__(
        self,
        emb_dim,
        emb_ks,
        emb_hs,
        n_freqs,
        hidden_channels,
        n_head=4,
        approx_qk_dim=512,
        activation="prelu",
        eps=1e-5,
        pool="mean",
        use_attention=False,
    ):
        super().__init__()
        bidirectional = False

        self.global_atten_causal = True

        self.pool = pool

        self.E = math.ceil(approx_qk_dim * 1.0 / n_freqs)  # approx_qk_dim is only approximate

        self.V_dim = emb_dim // n_head
        self.H = hidden_channels
        in_channels = emb_dim * emb_ks
        self.in_channels = in_channels
        self.n_freqs = n_freqs

        ## intra RNN can be optimized by conv or linear because the frequence length are not very large
        self.intra_norm = LayerNormalization4D_old(emb_dim, eps=eps)
        self.intra_rnn = nn.LSTM(in_channels, hidden_channels, 1, batch_first=True, bidirectional=True)
        self.intra_linear = nn.ConvTranspose1d(hidden_channels * 2, emb_dim, emb_ks, stride=emb_hs)
        self.emb_dim = emb_dim
        self.emb_ks = emb_ks
        self.emb_hs = emb_hs

        # inter RNN
        self.inter_norm = LayerNormalization4D_old(emb_dim, eps=eps)
        self.inter_rnn = nn.LSTM(in_channels, hidden_channels, 1, batch_first=True, bidirectional=bidirectional)
        self.inter_linear = nn.ConvTranspose1d(hidden_channels * (bidirectional + 1), emb_dim, emb_ks, stride=emb_hs)

        # attention
        self.use_attention = use_attention

        if self.use_attention:
            self.pool_atten_causal = Attention_STFT_causal(
                emb_dim=emb_dim,
                n_freqs=n_freqs,
                approx_qk_dim=approx_qk_dim,
                n_head=n_head,
                activation=activation,
                eps=eps,
            )

    def init_buffers(self, batch_size, device):
        return None

    # def _unfold_timedomain(self, x):
    #     BQ, C, T= x.shape
    #     # print("shape of x is {}".format(x.shape))
    #     # [117, 16, 4801] for causality testing
    #     # 4800 if training
    #     x = torch.split(x, self.lstm_fold_chunk, dim=-1) # [Num_chunk, BQ, C, 100]
    #     x = torch.cat(x, dim=0).reshape(-1, BQ, C, self.lstm_fold_chunk) # [Num_chunk, BQ, C, 100]
    #     x = x.permute(1, 0, 3, 2) # [BQ, Num_chunk, 100, C]
    #     return x

    def forward(self, x, init_state=None):
        """GridNetBlock Forward.

        Args:
            x: [B, C, T, Q]
            out: [B, C, T, Q]
        """
        B, C, old_T, old_Q = x.shape
        # print("shape of x is {}".format(x.shape))
        # print("old q is {}".format(old_Q))
        # print("dim just entered grid net block is {}".format(x.shape))
        # [1, 16, 4801, 117]
        T = math.ceil((old_T - self.emb_ks) / self.emb_hs) * self.emb_hs + self.emb_ks
        Q = math.ceil((old_Q - self.emb_ks) / self.emb_hs) * self.emb_hs + self.emb_ks
        x = F.pad(x, (0, Q - old_Q, 0, T - old_T))

        # ===========================Intra RNN start================================
        # define intra RNN
        input_ = x
        intra_rnn = self.intra_norm(input_)  # [B, C, T, Q]
        intra_rnn = intra_rnn.transpose(1, 2).contiguous().view(B * T, C, Q)  # [BT, C, Q]

        intra_rnn = torch.split(intra_rnn, self.emb_ks, dim=-1)  # [Q/I, BT, C, I]
        intra_rnn = torch.stack(intra_rnn, dim=0)
        intra_rnn = intra_rnn.permute(1, 2, 3, 0).flatten(1, 2)  # [BT, CI, Q/I]
        intra_rnn = intra_rnn.transpose(1, 2)  # [BT, -1, nC*emb_ks]
        self.intra_rnn.flatten_parameters()

        # apply intra frame LSTM
        intra_rnn, _ = self.intra_rnn(intra_rnn)  # [BT, -1, H]
        intra_rnn = intra_rnn.transpose(1, 2)  # [BT, H, -1]
        intra_rnn = self.intra_linear(intra_rnn)  # [BT, C, Q]
        intra_rnn = intra_rnn.view([B, T, C, Q])
        intra_rnn = intra_rnn.transpose(1, 2).contiguous()  # [B, C, T, Q]
        intra_rnn = intra_rnn + input_  # [B, C, T, Q]
        intra_rnn = intra_rnn[:, :, :, :old_Q]  # [B, C, T, Q]
        Q = old_Q

        # ===========================Intra RNN end================================

        # print("dim after intra rnn is {}".format(intra_rnn.shape))
        # [1, 16, 4801, 117]
        # [B, C, T, Q]

        # inter_rnn=intra_rnn
        # ===========================Inter RNN start================================
        # fold the time domain to chunk
        input_ = intra_rnn

        inter_rnn = self.inter_norm(intra_rnn)  # [B, C, T, Q]
        inter_rnn = inter_rnn.transpose(1, 3).reshape(B * Q, T, C)
        # inter_rnn = (
        #     inter_rnn.permute(0, 3, 1, 2).contiguous().view(B * Q, C, T)
        # )  # [BF, C, T]

        # print("dim of inter rnn is {}".format(inter_rnn.shape))
        # [117, 16, 4801]

        self.inter_rnn.flatten_parameters()
        # print("inter rnn shape is {}".format(inter_rnn.shape))
        # [133, 400, 16]
        inter_rnn, _ = self.inter_rnn(inter_rnn)  # [B * Q, -1, H]
        inter_rnn = inter_rnn.transpose(1, 2)  # [BF, H, -1]
        inter_rnn = self.inter_linear(inter_rnn)  # [BF, C, T]

        _, new_C, new_T = inter_rnn.shape
        inter_rnn = inter_rnn.reshape(B, Q, new_C, new_T)
        inter_rnn = inter_rnn.permute(0, 2, 3, 1)
        # print("shape of inter rnn is {}".format(inter_rnn.shape)) # [133, 16, 4800]
        # print("shape of input_ is {}".format(input_.shape)) # [1, 16, 4800, 133]
        inter_rnn = inter_rnn + input_
        # ===========================Inter RNN end================================

        # inter rnn shape is [B, C, T, Q]

        # ===========================attention start================================
        if self.use_attention:
            out = inter_rnn  # [B, C, T, Q]

            inter_rnn = inter_rnn.permute(0, 2, 3, 1)
            inter_rnn = self.pool_atten_causal(inter_rnn)  # B T Q C
            inter_rnn = inter_rnn.permute(0, 3, 1, 2)  # [B, C, T, Q]
            inter_rnn = out + inter_rnn  # B, C, T, Q

            # Output is inter_rnn by default
            # inter_rnn = inter_rnn.reshape(B, Q, T, C)
            # inter_rnn = inter_rnn.permute(0, 3, 2, 1) # B C T Q
            inter_rnn = inter_rnn[..., :old_T, :]
        # ===========================attention end================================

        # print("final output inter rnn dimension is {}".format(inter_rnn.shape))
        # print("old T is {}".format(old_T))

        # print("final output dimension is {}".format(inter_rnn.shape))
        # [2, 16, 4800, 133] [B, C, T, Q]

        #     return inter_rnn, init_state#, [t0 - t0_0, t1 - t0, t2 - t2_0, t3 - t2, t5 - t4, t7 - t6]
        # else:
        return inter_rnn, init_state