File size: 10,211 Bytes
0d80816
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# This module is from [WeNet](https://github.com/wenet-e2e/wenet).

# ## Citations

# ```bibtex
# @inproceedings{yao2021wenet,
#   title={WeNet: Production oriented Streaming and Non-streaming End-to-End Speech Recognition Toolkit},
#   author={Yao, Zhuoyuan and Wu, Di and Wang, Xiong and Zhang, Binbin and Yu, Fan and Yang, Chao and Peng, Zhendong and Chen, Xiaoyu and Xie, Lei and Lei, Xin},
#   booktitle={Proc. Interspeech},
#   year={2021},
#   address={Brno, Czech Republic },
#   organization={IEEE}
# }

# @article{zhang2022wenet,
#   title={WeNet 2.0: More Productive End-to-End Speech Recognition Toolkit},
#   author={Zhang, Binbin and Wu, Di and Peng, Zhendong and Song, Xingchen and Yao, Zhuoyuan and Lv, Hang and Xie, Lei and Yang, Chao and Pan, Fuping and Niu, Jianwei},
#   journal={arXiv preprint arXiv:2203.15455},
#   year={2022}
# }
#

from typing import Optional

import torch
from torch import nn
from modules.wenet_extractor.utils.mask import make_pad_mask


class Predictor(nn.Module):
    def __init__(
        self,
        idim,
        l_order,
        r_order,
        threshold=1.0,
        dropout=0.1,
        smooth_factor=1.0,
        noise_threshold=0,
        tail_threshold=0.45,
    ):
        super().__init__()

        self.pad = nn.ConstantPad1d((l_order, r_order), 0.0)
        self.cif_conv1d = nn.Conv1d(idim, idim, l_order + r_order + 1, groups=idim)
        self.cif_output = nn.Linear(idim, 1)
        self.dropout = torch.nn.Dropout(p=dropout)
        self.threshold = threshold
        self.smooth_factor = smooth_factor
        self.noise_threshold = noise_threshold
        self.tail_threshold = tail_threshold

    def forward(
        self,
        hidden,
        target_label: Optional[torch.Tensor] = None,
        mask: torch.Tensor = torch.tensor(0),
        ignore_id: int = -1,
        mask_chunk_predictor: Optional[torch.Tensor] = None,
        target_label_length: Optional[torch.Tensor] = None,
    ):
        h = hidden
        context = h.transpose(1, 2)
        queries = self.pad(context)
        memory = self.cif_conv1d(queries)
        output = memory + context
        output = self.dropout(output)
        output = output.transpose(1, 2)
        output = torch.relu(output)
        output = self.cif_output(output)
        alphas = torch.sigmoid(output)
        alphas = torch.nn.functional.relu(
            alphas * self.smooth_factor - self.noise_threshold
        )
        if mask is not None:
            mask = mask.transpose(-1, -2).float()
            alphas = alphas * mask
        if mask_chunk_predictor is not None:
            alphas = alphas * mask_chunk_predictor
        alphas = alphas.squeeze(-1)
        mask = mask.squeeze(-1)
        if target_label_length is not None:
            target_length = target_label_length
        elif target_label is not None:
            target_length = (target_label != ignore_id).float().sum(-1)
        else:
            target_length = None
        token_num = alphas.sum(-1)
        if target_length is not None:
            alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))
        elif self.tail_threshold > 0.0:
            hidden, alphas, token_num = self.tail_process_fn(
                hidden, alphas, token_num, mask=mask
            )

        acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)

        if target_length is None and self.tail_threshold > 0.0:
            token_num_int = torch.max(token_num).type(torch.int32).item()
            acoustic_embeds = acoustic_embeds[:, :token_num_int, :]

        return acoustic_embeds, token_num, alphas, cif_peak

    def tail_process_fn(
        self,
        hidden,
        alphas,
        token_num: Optional[torch.Tensor] = None,
        mask: Optional[torch.Tensor] = None,
    ):
        b, t, d = hidden.size()
        tail_threshold = self.tail_threshold
        if mask is not None:
            zeros_t = torch.zeros((b, 1), dtype=torch.float32, device=alphas.device)
            ones_t = torch.ones_like(zeros_t)
            mask_1 = torch.cat([mask, zeros_t], dim=1)
            mask_2 = torch.cat([ones_t, mask], dim=1)
            mask = mask_2 - mask_1
            tail_threshold = mask * tail_threshold
            alphas = torch.cat([alphas, zeros_t], dim=1)
            alphas = torch.add(alphas, tail_threshold)
        else:
            tail_threshold_tensor = torch.tensor(
                [tail_threshold], dtype=alphas.dtype
            ).to(alphas.device)
            tail_threshold_tensor = torch.reshape(tail_threshold_tensor, (1, 1))
            alphas = torch.cat([alphas, tail_threshold_tensor], dim=1)
        zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)
        hidden = torch.cat([hidden, zeros], dim=1)
        token_num = alphas.sum(dim=-1)
        token_num_floor = torch.floor(token_num)

        return hidden, alphas, token_num_floor

    def gen_frame_alignments(
        self, alphas: torch.Tensor = None, encoder_sequence_length: torch.Tensor = None
    ):
        batch_size, maximum_length = alphas.size()
        int_type = torch.int32

        is_training = self.training
        if is_training:
            token_num = torch.round(torch.sum(alphas, dim=1)).type(int_type)
        else:
            token_num = torch.floor(torch.sum(alphas, dim=1)).type(int_type)

        max_token_num = torch.max(token_num).item()

        alphas_cumsum = torch.cumsum(alphas, dim=1)
        alphas_cumsum = torch.floor(alphas_cumsum).type(int_type)
        alphas_cumsum = alphas_cumsum[:, None, :].repeat(1, max_token_num, 1)

        index = torch.ones([batch_size, max_token_num], dtype=int_type)
        index = torch.cumsum(index, dim=1)
        index = index[:, :, None].repeat(1, 1, maximum_length).to(alphas_cumsum.device)

        index_div = torch.floor(torch.true_divide(alphas_cumsum, index)).type(int_type)
        index_div_bool_zeros = index_div.eq(0)
        index_div_bool_zeros_count = torch.sum(index_div_bool_zeros, dim=-1) + 1
        index_div_bool_zeros_count = torch.clamp(
            index_div_bool_zeros_count, 0, encoder_sequence_length.max()
        )
        token_num_mask = (~make_pad_mask(token_num, max_len=max_token_num)).to(
            token_num.device
        )
        index_div_bool_zeros_count *= token_num_mask

        index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(
            1, 1, maximum_length
        )
        ones = torch.ones_like(index_div_bool_zeros_count_tile)
        zeros = torch.zeros_like(index_div_bool_zeros_count_tile)
        ones = torch.cumsum(ones, dim=2)
        cond = index_div_bool_zeros_count_tile == ones
        index_div_bool_zeros_count_tile = torch.where(cond, zeros, ones)

        index_div_bool_zeros_count_tile_bool = index_div_bool_zeros_count_tile.type(
            torch.bool
        )
        index_div_bool_zeros_count_tile = 1 - index_div_bool_zeros_count_tile_bool.type(
            int_type
        )
        index_div_bool_zeros_count_tile_out = torch.sum(
            index_div_bool_zeros_count_tile, dim=1
        )
        index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out.type(
            int_type
        )
        predictor_mask = (
            (
                ~make_pad_mask(
                    encoder_sequence_length, max_len=encoder_sequence_length.max()
                )
            )
            .type(int_type)
            .to(encoder_sequence_length.device)
        )
        index_div_bool_zeros_count_tile_out = (
            index_div_bool_zeros_count_tile_out * predictor_mask
        )

        predictor_alignments = index_div_bool_zeros_count_tile_out
        predictor_alignments_length = predictor_alignments.sum(-1).type(
            encoder_sequence_length.dtype
        )
        return predictor_alignments.detach(), predictor_alignments_length.detach()


class MAELoss(nn.Module):
    def __init__(self, normalize_length=False):
        super(MAELoss, self).__init__()
        self.normalize_length = normalize_length
        self.criterion = torch.nn.L1Loss(reduction="sum")

    def forward(self, token_length, pre_token_length):
        loss_token_normalizer = token_length.size(0)
        if self.normalize_length:
            loss_token_normalizer = token_length.sum().type(torch.float32)
        loss = self.criterion(token_length, pre_token_length)
        loss = loss / loss_token_normalizer
        return loss


def cif(hidden: torch.Tensor, alphas: torch.Tensor, threshold: float):
    batch_size, len_time, hidden_size = hidden.size()

    # loop varss
    integrate = torch.zeros([batch_size], device=hidden.device)
    frame = torch.zeros([batch_size, hidden_size], device=hidden.device)
    # intermediate vars along time
    list_fires = []
    list_frames = []

    for t in range(len_time):
        alpha = alphas[:, t]
        distribution_completion = (
            torch.ones([batch_size], device=hidden.device) - integrate
        )

        integrate += alpha
        list_fires.append(integrate)

        fire_place = integrate >= threshold
        integrate = torch.where(
            fire_place,
            integrate - torch.ones([batch_size], device=hidden.device),
            integrate,
        )
        cur = torch.where(fire_place, distribution_completion, alpha)
        remainds = alpha - cur

        frame += cur[:, None] * hidden[:, t, :]
        list_frames.append(frame)
        frame = torch.where(
            fire_place[:, None].repeat(1, hidden_size),
            remainds[:, None] * hidden[:, t, :],
            frame,
        )

    fires = torch.stack(list_fires, 1)
    frames = torch.stack(list_frames, 1)
    list_ls = []
    len_labels = torch.round(alphas.sum(-1)).int()
    max_label_len = len_labels.max()
    for b in range(batch_size):
        fire = fires[b, :]
        l = torch.index_select(
            frames[b, :, :], 0, torch.nonzero(fire >= threshold).squeeze()
        )
        pad_l = torch.zeros(
            [int(max_label_len - l.size(0)), hidden_size], device=hidden.device
        )
        list_ls.append(torch.cat([l, pad_l], 0))
    return torch.stack(list_ls, 0), fires