-
Notifications
You must be signed in to change notification settings - Fork 109
/
transducer_runners.py
133 lines (119 loc) · 5.48 KB
/
transducer_runners.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
import tensorflow as tf
import tensorflow.keras.mixed_precision.experimental as mixed_precision
from trainer.base_runners import BaseTrainer
from losses.rnnt_losses import USE_TF,tf_rnnt_loss,rnnt_loss
from AMmodel.transducer_wrap import Transducer
from utils.text_featurizers import TextFeaturizer
import logging
class TransducerTrainer(BaseTrainer):
def __init__(self,
speech_featurizer,
text_featurizer: TextFeaturizer,
config: dict,
is_mixed_precision: bool = False,
strategy=None
):
"""
Args:
config: the 'running_config' part in YAML config file'
text_featurizer: the TextFeaturizer instance
is_mixed_precision: a boolean for using mixed precision or not
"""
super(TransducerTrainer, self).__init__(config)
self.speech_featurizer=speech_featurizer
self.text_featurizer = text_featurizer
self.is_mixed_precision = is_mixed_precision
self.set_strategy(strategy)
if USE_TF:
self.rnnt_loss=tf_rnnt_loss
else:
self.rnnt_loss=rnnt_loss
def set_train_metrics(self):
self.train_metrics = {
"transducer_loss": tf.keras.metrics.Mean("train_transducer_loss", dtype=tf.float32)
}
def set_eval_metrics(self):
self.eval_metrics = {
"transducer_loss": tf.keras.metrics.Mean("eval_transducer_loss", dtype=tf.float32)
}
@tf.function(experimental_relax_shapes=True)
def _train_step(self, batch):
features, wavs, input_length, labels, label_length = batch
pred_inp=labels
target=labels[:,1:]
label_length-=1
with tf.GradientTape() as tape:
if self.model.mel_layer is not None:
logits = self.model([wavs, pred_inp], training=True)
else:
logits = self.model([features, pred_inp], training=True)
tape.watch(logits)
# print(logits.shape,target.shape)
if USE_TF:
per_train_loss=self.rnnt_loss(logits=logits, labels=target
, label_length=label_length, logit_length=input_length)
else:
per_train_loss = self.rnnt_loss(
logits=logits, labels=labels, label_length=label_length,
logit_length=(input_length // self.model.time_reduction_factor),
blank=self.text_featurizer.blank)
train_loss = tf.nn.compute_average_loss(per_train_loss,
global_batch_size=self.global_batch_size)
if self.is_mixed_precision:
scaled_train_loss = self.optimizer.get_scaled_loss(train_loss)
if self.is_mixed_precision:
scaled_gradients = tape.gradient(scaled_train_loss, self.model.trainable_variables)
gradients = self.optimizer.get_unscaled_gradients(scaled_gradients)
else:
gradients = tape.gradient(train_loss, self.model.trainable_variables)
self.optimizer.apply_gradients(zip(gradients, self.model.trainable_variables))
self.train_metrics["transducer_loss"].update_state(per_train_loss)
@tf.function(experimental_relax_shapes=True)
def _eval_step(self, batch):
features, wavs, input_length, labels, label_length = batch
pred_inp = labels
target = labels[:, 1:]
label_length -= 1
if self.model.mel_layer is not None:
logits=self.model([wavs, pred_inp], training=False)
else:
logits = self.model([features, pred_inp], training=False)
if USE_TF:
eval_loss = self.rnnt_loss(logits=logits, labels=target
, label_length=label_length,
logit_length=input_length,
)
else:
eval_loss = self.rnnt_loss(
logits=logits, labels=target, label_length=label_length,
logit_length=(input_length // self.model.time_reduction_factor),
blank=self.text_featurizer.blank)
self.eval_metrics["transducer_loss"].update_state(eval_loss)
def compile(self,
model: Transducer,
optimizer: any,
max_to_keep: int = 10):
f, c = self.speech_featurizer.compute_feature_dim()
with self.strategy.scope():
self.model = model
if self.model.mel_layer is not None:
self.model._build([1, 16000, 1])
else:
self.model._build([1, 80, f, c])
self.model.summary(line_length=100)
try:
self.load_checkpoint()
except:
logging.info('trainer resume failed')
self.optimizer = tf.keras.optimizers.get(optimizer)
if self.is_mixed_precision:
self.optimizer = mixed_precision.LossScaleOptimizer(self.optimizer, "dynamic")
self.set_progbar()
# self.load_checkpoint()
def fit(self, epoch=None):
if epoch is not None:
self.epochs=epoch
self.train_progbar.set_description_str(
f"[Train] [Epoch {epoch}/{self.config['num_epochs']}]")
self._train_batches()
self._check_eval_interval()