-
Notifications
You must be signed in to change notification settings - Fork 2
/
model.py
406 lines (365 loc) · 13.3 KB
/
model.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
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
from typing import Tuple
from torch import Tensor
import torch
import torch.nn as nn
class PredNet(nn.Module):
"""Implements the functionalities of the
predict network in the model architecture
Attributes
----------
emb : nn.Module
The network's embedding layer
lstm : nn.Module
The network's RNN layer
"""
def __init__(
self,
vocab_size: int,
emb_dim: int,
pad_idx: int,
hidden_size: int,
n_layers: int,
dropout: float
) -> None:
"""Constructs all the necessary attributes
Args:
vocab_size (int): The number of vocabularies in the dataset,
used for the embedding layer
emb_dim (int): The embedding layer dimensionality
pad_idx (int): the padding index
hidden_size (int): The RNN's hidden layer size
n_layers (int): the number of stacked LSTM layers in the network
dropout (float): the dropout rate for each RNN layer
"""
super().__init__()
self.n_layers = n_layers
self.hidden_size = hidden_size
self.emb = nn.Embedding(
vocab_size, emb_dim, padding_idx=pad_idx
)
self.lstm = nn.LSTM(
input_size=emb_dim,
hidden_size=hidden_size,
num_layers=n_layers,
batch_first=True,
dropout=dropout
)
def forward(
self,
x: Tensor,
hn: Tensor,
cn: Tensor
) -> Tuple[Tensor, Tensor, Tensor]:
self.validate_dims(x, hn, cn)
out = self.emb(x)
out, (hn, cn) = self.lstm(out, (hn, cn))
return out, hn, cn
def get_zeros_hidden_state(self, batch_size: int) -> Tuple[Tensor, Tensor]:
return (
torch.zeros((self.n_layers, batch_size, self.hidden_size)),
torch.zeros((self.n_layers, batch_size, self.hidden_size))
)
def validate_dims(
self,
x: Tensor,
hn: Tensor,
cn: Tensor
) -> None:
assert hn.shape[0] == self.n_layers, \
'The hidden state should match the number of layers'
assert hn.shape[2] == self.hidden_size, \
'The hidden state should match the hiden size'
assert cn.shape[0] == self.n_layers, \
'The cell state should match the number of layers'
assert cn.shape[2] == self.hidden_size, \
'The cell state should match the hiden size'
class TransNet(nn.Module):
"""Implements the functionalities of the
transcription network in the model architecture, where
the input is the speech features and projects it to high
level feature representation.
Attributes
----------
lstm : nn.Module
The network's RNN layer
"""
def __init__(
self,
input_size: int,
hidden_size: int,
n_layers: int,
dropout: float,
is_bidirectional: bool
) -> None:
"""
Args:
input_size (int): The number of input features per time step,
hidden_size (int): The RNN's hidden layer size
n_layers (int): the number of stacked LSTM layers in the network
dropout (float): the dropout rate for each RNN layer
is_bidirectional (bool): if the RNN layers are bidirectional or not
"""
super().__init__()
self.lstm = nn.LSTM(
input_size=input_size,
hidden_size=hidden_size,
num_layers=n_layers,
batch_first=True,
dropout=dropout,
bidirectional=is_bidirectional
)
def forward(self, x: Tensor) -> Tensor:
out, *_ = self.lstm(x)
return out
class JoinNet(nn.Module):
"""Implements the functionalities of the
Join network in the model architecture, where
the inputs are high speech features at time tn and the prediction at step u
and predicts the next character or phi based on that, there are two
moods of operations additive and multiplicative mood.
Attributes
----------
MOODES: dict
Maps the mode to a function
fc : nn.Module
The network's fully connected layer that maps the
features into vocabulary distribution
join_mood: Callable
The mood of operation
"""
MODES = {
'multiplicative': lambda f, g: f * g,
'mul': lambda f, g: f * g,
'additive': lambda f, g: f + g,
'add': lambda f, g: f + g
}
def __init__(
self,
input_size: int,
vocab_size: int,
mode: str
) -> None:
"""
Args:
input_size (int): The dimension of each timesteps features
vocab_size (int): The number of vocabulary in the corpus
mode (str): The mode of operations, either mul
for multiplicative or add for additive
"""
super().__init__()
self.join_mood = self.MODES[mode]
self.fc = nn.Linear(
in_features=input_size,
out_features=vocab_size
)
def forward(self, f: Tensor, g: Tensor) -> Tensor:
"""performs forward propagation step
Args:
f (Tensor): The transcription vector at time t of shape (B, 1, h)
g (Tensor): The prediction vector at step u of shape (B, 1, h)
Returns:
Tensor: vocabulary distribution
"""
out = self.join_mood(f, g)
out = self.fc(out)
return torch.softmax(out, dim=-1)
class Model(nn.Module):
"""Implements the full RNN-T model which consists of
- prediction network
- transcription network
- join network
Attributes
----------
prednet : nn.Module
The prediction network
transnet: nn.Module
The transcrption network
joinnet: nn.Module
The join network
device: str
The device to do the operations on.
default to cuda.
phi_idx: int
The index of phi symbol
pad_idx: int
The index of the padding symbol
sos_idx: int
The index of the start of sentence symbol
"""
def __init__(
self,
prednet_params: dict,
transnet_params: dict,
joinnet_params: dict,
phi_idx: int,
pad_idx: int,
sos_idx: int,
device='cuda'
) -> None:
super().__init__()
self.prednet = PredNet(**prednet_params).to(device)
self.transnet = TransNet(**transnet_params).to(device)
self.joinnet = JoinNet(**joinnet_params).to(device)
self.prednet_hidden_size = prednet_params['hidden_size']
self.device = device
self.phi_idx = phi_idx
self.pad_idx = pad_idx
self.sos_idx = sos_idx
def forward(
self,
x: Tensor,
max_length: int
) -> Tensor:
# TODO: Code Refactoring and documentation
# FIX: feeding phis to the predict network
batch_size, T, *_ = x.shape
counter = self.get_counter_start(batch_size, T)
counter_ceil = self.get_counter_ceil(counter, T)
term_state = torch.zeros(batch_size)
trans_result = self.feed_into_transnet(x)
# reshaping the results (B, T, F) -> (B * T, F)
trans_result = trans_result.reshape(batch_size * T, -1)
h, c = self.prednet.get_zeros_hidden_state(batch_size)
h = h.to(self.device)
c = c.to(self.device)
gu = self.get_sos_seed(batch_size)
t = 0
while True:
t += 1
preds, h, c = self.predict_next(gu, h, c, counter, trans_result)
if t == 1:
result = preds
else:
result = torch.concat([result, preds], dim=1)
gu = self.keep_last_char(gu, torch.argmax(preds, dim=-1))
counter, update_mask, term_state = self.update_states(
gu, counter, counter_ceil, term_state, t
)
if (update_mask.sum().item() == batch_size) or (max_length == t):
break
return result, term_state
def keep_last_char(self, gu: Tensor, preds: Tensor) -> Tensor:
"""Keeps the characters only in the gu Tensors, where the phis get
neglected to not be passed to the prednet.
Args:
gu (Tensor): The u-1 predicted chars.
preds (Tensor): The latest predicted chars.
Returns:
Tensor: Updated gu where the character are kept, and the phis in
preds are replaced with the chars from gu.
"""
is_phi = preds == self.phi_idx
return (is_phi * gu) + (~is_phi * preds)
def update_states(
self,
gu: Tensor,
counter: Tensor,
counter_ceil: Tensor,
term_state: Tensor,
t: int
) -> Tuple[Tensor, Tensor, Tensor]:
"""Updates the positional related tensors, these positionals store
the state of the pointers and these are teh counter and the term_state
tensors
Args:
gu (Tensor): The latest predicted characters
counter (Tensor): The counter tensor
counter_ceil (Tensor): the counter ceil that store the limit of
the pointers.
term_state (Tensor): The terminate state that stores where the
results end.
t (int): The latest time step
Returns:
Tuple[Tensor, Tensor, Tensor]: A tuple of the updated states
"""
counter = counter + ((gu.cpu() == self.phi_idx).squeeze())
counter, update_mask = self.clip_counter(counter, counter_ceil)
term_state = self.update_termination_state(
term_state, update_mask, t
)
return counter, update_mask, term_state
def predict_next(
self,
gu: Tensor,
h: Tensor,
c: Tensor,
counter: Tensor,
trans_result: Tensor
) -> Tuple[Tensor, Tensor, Tensor]:
"""Does a single prediction over time
Args:
gu (Tensor): The latest character predicted
h (Tensor): The latest hidden states out of
the prediction network
c (Tensor): The latest cell state out of the prediction network
counter (Tensor): The counter vector that tracks the pointers over
the time frames
trans_result (Tensor): The frames out of the Transcription network
Returns:
Tuple[Tensor, Tensor, Tensor]: A tuple of the prediction, hidden
state and cell state
"""
out, h, c = self.prednet(gu, h, c)
fy = trans_result[counter, :].unsqueeze(dim=1)
preds = self.joinnet(fy, out)
return preds, h, c
def get_counter_ceil(
self, counter: Tensor, T: int
) -> Tensor:
return counter + T - 1
def get_sos_seed(
self, batch_size: int
) -> Tensor:
return torch.LongTensor([[self.sos_idx]] * batch_size).to(self.device)
def feed_into_transnet(self, x: Tensor) -> Tensor:
return self.transnet(x)
def feed_into_prednet(
self, yu: Tensor, h: Tensor, c: Tensor
) -> Tuple[Tensor, Tensor, Tensor]:
return self.transnet(yu, h, c)
def get_counter_start(
self, batch_size: int, max_size: int
) -> Tensor:
return torch.arange(0, batch_size * max_size, max_size)
def clip_counter(
self, counter: Tensor, ceil_vector: Tensor
) -> Tuple[Tensor, Tensor]:
"""Clips the counter to the ceil values,
if the value at index i in the counter
exceeded teh value at index i at the ceil_vector
it will be assigned to the ceil_vector[i]
Args:
counter (Tensor): The counter vector to be updated
ceil_vector (Tensor): The maximum value at each index of the
counter values
Returns:
Tuple[Tensor, Tensor]: A tuple of the updated counter
and a boolean tensor where indicates where the values
are exceeded the limit
"""
update_mask = counter >= ceil_vector
upper_bounded = update_mask * ceil_vector
kept_counts = (counter < ceil_vector) * counter
return upper_bounded + kept_counts, update_mask
def update_termination_state(
self,
term_state: Tensor,
update_mask: Tensor,
last_index: int
) -> Tensor:
"""Updates the termination state, where the
it stores if an example m reached the end of transcription or not
Args:
term_state (Tensor): The latest termination state of (N,) shape
where the value at index i eitehr 0 or number if 0 it's not
terminated yet otherwise the number indicates the latest position
to consider.
update_mask (Tensor): The update mask tensor resulted from the
clip operation.
last_index (int): The last index reached in the iteration loop.
Returns:
Tensor: The updated term_state tensor
"""
is_unended = term_state == 0
to_update = is_unended & update_mask
return term_state + to_update * last_index