-
Notifications
You must be signed in to change notification settings - Fork 1
/
transducer.py
287 lines (258 loc) · 11.1 KB
/
transducer.py
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
import torch
import torchaudio
import torch.nn.functional as F
from torch import nn
from transformers.file_utils import ModelOutput
from dataclasses import dataclass
from typing import Optional, Tuple, List
from .loss import LatticeLoss
from .utils import add_blank, get_padding_mask
from .greedy_search import rnnt_greedy_search, frnnt_greedy_search
@dataclass
class ClsOutput(ModelOutput):
loss: torch.FloatTensor = None
loss_rnnt: Optional[torch.FloatTensor] = None
head_logits: Optional[torch.FloatTensor] = None
class ENT(nn.Module):
def __init__(self,
vocab_size,
encoder,
predictor,
joint,
joint_emotion=None,
head_emotion=None,
joint_weight=0.0,
head_weight=0.5,
lm_weight=0.0,
head_lm=None,
rnnt_weight=0.5,
blank=0,
ignore_id=-1):
super().__init__()
self.blank = blank
self.ignore_id = ignore_id
self.vocab_size = vocab_size
self.sos = vocab_size - 1
self.eos = vocab_size - 1
self.encoder = encoder
self.predictor = predictor
self.joint = joint
self.head_weight = head_weight
self.joint_weight = joint_weight
self.lm_weight = lm_weight
self.rnnt_weight = rnnt_weight
self.use_emotion = True if joint_weight or head_weight else False
self.finetune = self.encoder.finetune
if self.joint_weight:
self.joint_emotion = joint_emotion
self.joint_emofn = LatticeLoss()
if self.head_weight:
self.head_emotion = head_emotion
if self.lm_weight:
self.head_lm = head_lm
def forward(self, audio, audio_length, text, text_length, label):
"""
Args:
audio: (Batch, Length, ...)
audio_length: (Batch, )
text: (Batch, Length)
text_length: (Batch,)
"""
# encoder
enc_out, _ = self.encoder(audio, audio_length)
# predictor
ys_in_pad = add_blank(text, self.blank, self.ignore_id)
pred_out = self.predictor(ys_in_pad)
# joint
joint_out = self.joint(enc_out, pred_out) # [B,T,U,V]
padded_text_length = torch.add(text_length, 1)
if self.head_weight:
head_logits = self.head_emotion(enc_out, audio_length, pred_out, padded_text_length)
loss_emohead = self.head_weight * F.cross_entropy(head_logits, label)
else:
head_logits = None
loss_emohead = 0
if self.joint_weight:
joint_logits = self.joint_emotion(enc_out, pred_out)
B, T, U, _ = joint_logits.shape
joint_label = label.view(B, 1, 1).expand(-1, T, U) # [B, T, U]
enc_mask = get_padding_mask(T, B, audio_length).unsqueeze(-1).expand(-1, -1, U).bool() # [B, T, U]
pred_mask = get_padding_mask(U, B, padded_text_length).unsqueeze(1).expand(-1, T, -1).bool() # [B, T, U]
mask = torch.logical_and(enc_mask, pred_mask)
# loss_emojoint = self.joint_weight * F.cross_entropy(joint_logits[mask], joint_label[mask])
loss_emojoint = self.joint_weight * self.joint_emofn(joint_logits, joint_label, mask)
else:
joint_logits = None
loss_emojoint = 0
if self.training and self.lm_weight:
lm_logits = self.head_lm(pred_out)
B, U, V = lm_logits.shape
mask = get_padding_mask(U, B, padded_text_length)
padded_text = F.pad(text, [0, 1], "constant", self.ignore_id)
padded_text = torch.scatter(padded_text, 1, text_length.unsqueeze(-1), self.eos)
loss_lm = self.lm_weight * F.cross_entropy(lm_logits[mask], padded_text[mask])
else:
loss_lm = 0
# rnnt-loss
if self.training:
rnnt_text = torch.where(text == self.ignore_id, 0, text).to(torch.int32)
rnnt_text_length = text_length.to(torch.int32)
audio_length = audio_length.to(torch.int32)
loss_rnnt = self.rnnt_weight * torchaudio.functional.rnnt_loss(
joint_out, rnnt_text, audio_length, rnnt_text_length, blank=self.blank, reduction="mean")
else:
loss_rnnt = 0
return ClsOutput(loss=loss_rnnt + loss_emohead + loss_emojoint + loss_lm, head_logits=head_logits)
def greedy_search(
self,
audio: torch.Tensor,
audio_length: torch.Tensor,
decoding_chunk_size: int = -1,
num_decoding_left_chunks: int = -1,
simulate_streaming: bool = False,
n_steps: int = 32,
):
""" greedy search
Args:
audio (torch.Tensor): (batch=1, max_len, feat_dim)
audio_length (torch.Tensor): (batch, )
decoding_chunk_size (int): decoding chunk for dynamic chunk
trained model.
<0: for decoding, use full chunk.
>0: for decoding, use fixed chunk size as set.
0: used for training, it's prohibited here
simulate_streaming (bool): whether do encoder forward in a
streaming fashion
Returns:
List[List[int]]: best path result
"""
_ = simulate_streaming
enc_out, _ = self.encoder(audio, audio_length)
hyps = rnnt_greedy_search(self, enc_out, audio_length, n_steps=n_steps)
# hyps = rnnt_beam_search(self, enc_out, audio_length, n_steps=n_steps)
return hyps
class FENT(nn.Module):
def __init__(self,
vocab_size,
encoder,
predictor_blank,
joint_blank,
predictor_vocab,
joint_vocab,
joint_emotion=None,
joint_weight=0.5,
head_emotion=None,
head_weight=0.5,
lm_weight=0.0,
head_lm=None,
rnnt_weight=0.5,
blank=0,
ignore_id=-1):
super().__init__()
self.blank = blank
self.ignore_id = ignore_id
self.vocab_size = vocab_size
self.sos = vocab_size - 1
self.eos = vocab_size - 1
self.encoder = encoder
self.predictor_blank = predictor_blank
self.joint_blank = joint_blank
self.predictor_vocab = predictor_vocab
self.joint_vocab = joint_vocab
self.head_weight = head_weight
self.joint_weight = joint_weight
self.lm_weight = lm_weight
self.rnnt_weight = rnnt_weight
self.use_emotion = True if joint_weight or head_weight else False
self.finetune = self.encoder.finetune
if self.joint_weight:
self.joint_emotion = joint_emotion
self.joint_emofn = LatticeLoss()
if self.head_weight:
self.head_emotion = head_emotion
if self.lm_weight:
self.head_lm = head_lm
def forward(self, audio, audio_length, text, text_length, label):
"""
Args:
audio: (Batch, Length, ...)
audio_length: (Batch, )
text: (Batch, Length)
text_length: (Batch,)
"""
# encoder
enc_out, enc_emo_out = self.encoder(audio, audio_length)
# predictor
ys_in_pad = add_blank(text, self.blank, self.ignore_id)
predb_out = self.predictor_blank(ys_in_pad)
predv_out = self.predictor_vocab(ys_in_pad)
# joint
jointb_out = self.joint_blank(enc_out if enc_emo_out is None else enc_emo_out, predb_out) # [B,T,U,1]
jointv_out = self.joint_vocab(enc_out, predv_out) # [B,T,U,V-1]
joint_out = torch.cat((jointb_out, jointv_out), dim=-1)
padded_text_length = torch.add(text_length, 1)
if self.head_weight:
head_logits = self.head_emotion(enc_out if enc_emo_out is None else enc_emo_out, audio_length, predb_out, padded_text_length)
loss_emohead = self.head_weight * F.cross_entropy(head_logits, label)
else:
head_logits = None
loss_emohead = 0
if self.joint_weight:
joint_logits = self.joint_emotion(enc_out if enc_emo_out is None else enc_emo_out, predb_out) # [B,T,U,4]
B, T, U, _ = joint_logits.shape
joint_label = label.view(B, 1, 1).expand(-1, T, U) # [B, T, U]
enc_mask = get_padding_mask(T, B, audio_length).unsqueeze(-1).expand(-1, -1, U).bool() # [B, T, U]
pred_mask = get_padding_mask(U, B, padded_text_length).unsqueeze(1).expand(-1, T, -1).bool() # [B, T, U]
mask = torch.logical_and(enc_mask, pred_mask)
# loss_emojoint = self.joint_weight * F.cross_entropy(joint_logits[mask], joint_label[mask])
loss_emojoint = self.joint_weight * self.joint_emofn(joint_logits, joint_label, mask)
else:
joint_logits = None
loss_emojoint = 0
if self.training and self.lm_weight:
lm_logits = self.head_lm(predv_out)
B, U, V = lm_logits.shape
mask = get_padding_mask(U, B, padded_text_length)
padded_text = F.pad(text, [0, 1], "constant", self.ignore_id)
padded_text = torch.scatter(padded_text, 1, text_length.unsqueeze(-1), self.eos)
loss_lm = self.lm_weight * F.cross_entropy(lm_logits[mask], padded_text[mask])
else:
loss_lm = 0
# rnnt-loss
if self.training:
rnnt_text = torch.where(text == self.ignore_id, 0, text).to(torch.int32)
rnnt_text_length = text_length.to(torch.int32)
audio_length = audio_length.to(torch.int32)
loss_rnnt = self.rnnt_weight * torchaudio.functional.rnnt_loss(
joint_out, rnnt_text, audio_length, rnnt_text_length, blank=self.blank, reduction="mean")
else:
loss_rnnt = 0
return ClsOutput(loss=loss_rnnt + loss_emohead + loss_emojoint + loss_lm, head_logits=head_logits)
def greedy_search(
self,
audio: torch.Tensor,
audio_length: torch.Tensor,
decoding_chunk_size: int = -1,
num_decoding_left_chunks: int = -1,
simulate_streaming: bool = False,
n_steps: int = 32,
):
""" greedy search
Args:
audio (torch.Tensor): (batch=1, max_len, feat_dim)
audio_length (torch.Tensor): (batch, )
decoding_chunk_size (int): decoding chunk for dynamic chunk
trained model.
<0: for decoding, use full chunk.
>0: for decoding, use fixed chunk size as set.
0: used for training, it's prohibited here
simulate_streaming (bool): whether do encoder forward in a
streaming fashion
Returns:
List[List[int]]: best path result
"""
_ = simulate_streaming
enc_out, enc_emo_out = self.encoder(audio, audio_length)
hyps = frnnt_greedy_search(self, enc_out, audio_length, enc_emo_out, n_steps=n_steps)
# hyps = frnnt_beam_search(self, enc_out, audio_length, enc_emo_out, n_steps=n_steps)
return hyps