-
Notifications
You must be signed in to change notification settings - Fork 17
/
model.py
223 lines (176 loc) · 7.15 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
#!/usr/bin/env python3
# encoding: utf-8
import torch
from tt.utils.net_utils import get_subsample, make_non_pad_mask
from tt.rnn.encoders import RNNEncoder
from tt.transducer.loss import TransLoss
from tt.transducer.rnn_decoder import RNNDecoder
from tt.transducer.transformer_decoder import TDecoder
from tt.transducer.transformer_encoder import TEncoder
from tt.transducer.utils import prepare_loss_inputs
from tt.transformer.mask import target_mask
from pytorch_lightning import LightningModule
from dataclasses import asdict
from tt.beam_search_transducer import BeamSearchTransducer
class Transducer(LightningModule):
def __init__(self, idim, odim, args, blank_id=0, training=True):
super(Transducer, self).__init__()
if 'lstm' in args.etype:
self.subsample = get_subsample(args, mode="asr", arch="rnn-t")
self.Encoder = RNNEncoder(
args.etype,
args.elayers,
args.eunits,
args.eprojs,
idim,
self.subsample,
args.dropout_rate
)
else:
self.Encoder = TEncoder(
idim,
args.enc_block_arch,
input_layer=args.transformer_enc_input_layer,
repeat_block=args.enc_block_repeat,
self_attn_type=args.transformer_enc_self_attn_type,
positional_encoding_type=args.transformer_enc_positional_encoding_type,
positionwise_activation_type=args.transformer_enc_pw_activation_type,
conv_mod_activation_type=args.transformer_enc_conv_mod_activation_type,
)
encoder_out = self.Encoder.enc_out
if 'lstm' in args.dec_type:
self.Decoder = RNNDecoder(
args.eprojs,
odim,
args.dec_type,
args.dlayers,
args.dunits,
blank_id,
args.dec_embed_dim,
args.joint_dim,
args.joint_activation_type,
args.dropout_rate_decoder,
args.dropout_rate_embed_decoder,
)
else:
self.Decoder = TDecoder(
odim,
encoder_out,
args.joint_dim,
args.dec_block_arch,
input_layer=args.transformer_dec_input_layer,
repeat_block=args.dec_block_repeat,
joint_activation_type=args.joint_activation_type,
positionwise_activation_type=args.transformer_dec_pw_activation_type,
dropout_rate_embed=args.dropout_rate_embed_decoder,
)
self.etype = args.etype
self.dec_type = args.dec_type
self.blank_id = blank_id
self.blank = args.sym_blank
if training:
self.criterion = TransLoss(args.trans_type, self.blank_id)
self.error_calculator = None
self.rnnlm = None
self.args = args
self.idim = idim
self.odim = odim
self.beamsearch = BeamSearchTransducer(
decoder=self.Decoder,
beam_size=args.beam_size,
lm=self.rnnlm,
lm_weight=args.lm_weight,
search_type="default",
score_norm=args.score_norm,
)
def forward(self, xs_pad, ilens, ys_pad):
xs_pad = xs_pad[:, : max(ilens)]
masks = make_non_pad_mask(ilens.tolist()).to(xs_pad.device).unsqueeze(-2)
if self.args.right_mask > -1:
masks = target_mask(masks, self.blank_id, self.args.right_mask)
# Encoder
if "lstm" in self.etype:
hs_pad, hs_mask, _ = self.Encoder(xs_pad, ilens)
else:
hs_pad, hs_mask = self.Encoder(xs_pad, masks)
# Loss input
ys_in_pad, target, pred_len, target_len = prepare_loss_inputs(ys_pad, hs_mask)
# Decoder
if "lstm" in self.dec_type:
pred_pad = self.Decoder(hs_pad, ys_in_pad)
else:
ys_mask = target_mask(ys_in_pad, self.blank_id)
pred_pad, _ = self.Decoder(ys_in_pad, ys_mask, hs_pad)
return pred_pad, target, pred_len, target_len
def configure_optimizers(self):
if self.args.opt == "adadelta":
optimizer = torch.optim.Adadelta(
self.parameters(), rho=0.95, eps=self.args.eps, weight_decay=self.args.weight_decay
)
elif self.args.opt == "adam":
optimizer = torch.optim.Adam(self.parameters(), weight_decay=self.args.weight_decay)
elif self.args.opt == "noam":
from tt.noam import get_std_opt
# For transformer-transducer, adim declaration is within the block definition.
# Thus, we need retrieve the most dominant value (d_hidden) for Noam scheduler.
adim = self.args.adim
optimizer = get_std_opt(
self.parameters(), adim, self.args.transformer_warmup_steps, self.args.transformer_lr
)
return optimizer
def training_step(self, batch, batch_idx):
xs_pad, ilens, ys_pad = batch
pred_pad, target, pred_len, target_len = self.forward(xs_pad, ilens, ys_pad)
loss = self.criterion(pred_pad, target, pred_len, target_len)
self.log('loss', loss)
return loss
def validation_step(self, batch, batch_idx):
xs_pad, ilens, ys_pad = batch
pred_pad, target, pred_len, target_len = self.forward(xs_pad, ilens, ys_pad)
val_loss = self.criterion(pred_pad, target, pred_len, target_len)
return self.log('val_loss', val_loss)
def encode_transformer(self, x):
"""Encode acoustic features.
Args:
x (ndarray): input acoustic feature (T, D)
Returns:
x (torch.Tensor): encoded features (T, attention_dim)
"""
self.eval()
x = torch.as_tensor(x).unsqueeze(0)
ret = torch.ones(x.size(1), x.size(1), dtype=torch.uint8)
mask = torch.tril(ret, diagonal=self.args.right_mask).unsqueeze(0)
enc_output, _ = self.Encoder(x, mask)
return enc_output.squeeze(0)
def encode_rnn(self, x):
"""Encode acoustic features.
Args:
x (ndarray): input acoustic feature (T, D)
Returns:
x (torch.Tensor): encoded features (T, attention_dim)
"""
self.eval()
ilens = [x.shape[0]]
x = x[:: self.subsample[0], :]
p = next(self.parameters())
h = torch.as_tensor(x, device=p.device, dtype=p.dtype)
hs = h.contiguous().unsqueeze(0)
hs, _, _ = self.Encoder(hs, ilens)
return hs.squeeze(0)
def recog(self, x):
"""Recognize input features.
Args:
x (ndarray): input acoustic feature (T, D)
beam_search (class): beam search class
Returns:
nbest_hyps (list): n-best decoding results
"""
if "transformer" in self.etype:
h = self.encode_transformer(x)
else:
h = self.encode_rnn(x)
nbest_hyps = self.beamsearch(h)
if isinstance(nbest_hyps, list):
return [asdict(n) for n in nbest_hyps]
else:
return asdict(nbest_hyps)