forked from roman-vygon/triplet_loss_kws
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathTripletClassifier.py
196 lines (159 loc) · 6.55 KB
/
TripletClassifier.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
import math
import os
from functools import partial
from ruamel.yaml import YAML
import nemo
import nemo.collections.asr as nemo_asr
from nemo.utils.lr_policies import CosineAnnealing
from nemo.collections.asr.helpers import (
monitor_classification_training_progress,
process_classification_evaluation_batch,
process_classification_evaluation_epoch,
)
from models.resnet import Res15, Res8
import argparse
logging = nemo.logging
from models.classifier import ClassificationNet
from layers.l2 import L2Regularizer
import json
from models.fc import LinearLayer
parser = argparse.ArgumentParser(description='Triplet loss classifier')
parser.add_argument('--gpu', type=int, help='gpu#', default=0)
parser.add_argument('--name', type=str, help='logdir name', default='test')
parser.add_argument('--enc_name', type=str, help='name of encoder logdir name', default='test')
parser.add_argument('--enc_step', type=int, help='encoder checkpoint step', default=0)
parser.add_argument('--num_classes', type=int, help='number of classes', default=35)
parser.add_argument('--hidden_size', type=int, help='size of hidden layers', default=64)
parser.add_argument('--manifest', type=str, help='manifest', default='10')
parser.add_argument('--model', type=str, help='model', default='Res8')
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
manifests = json.load(open('manifests.json', 'r'))
data_dir = '.'
background_dataset = data_dir + '/google_dataset_v2/google_speech_recognition_v2/background_manifest.json'
train_dataset = manifests[args.manifest]['train']
val_dataset = manifests[args.manifest]['dev']
yaml = YAML(typ="safe")
with open(f"configs/words{args.manifest}.yaml") as f:
jasper_params = yaml.load(f)
labels = jasper_params['labels']
sample_rate = jasper_params['sample_rate']
lr = 1e-3
num_epochs = 2
batch_size = 128
weight_decay = 0.001
num_classes = len(labels)
neural_factory = nemo.core.NeuralModuleFactory(
log_dir=data_dir + '/runs/' + args.name,
create_tb_writer=True)
tb_writer = neural_factory.tb_writer
train_data_layer = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=train_dataset,
sample_rate=sample_rate,
labels=labels,
batch_size=batch_size,
num_workers=0,
shuffle=True,
)
eval_data_layer = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=val_dataset,
sample_rate=sample_rate,
labels=labels,
batch_size=batch_size,
num_workers=0,
shuffle=True,
)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=sample_rate, **jasper_params["AudioToMelSpectrogramPreprocessor"],
)
assert args.model in ['Res8', 'Res15', 'Quartz']
if args.model == 'Res8':
encoder = Res8(args.hidden_size).to('cuda')
encoder.restore_from(f'./runs/{args.enc_name}/checkpoints/Res8-STEP-{str(args.enc_step)}.pt')
elif args.model == 'Res15':
encoder = Res15(args.hidden_size).to('cuda')
encoder.restore_from(f'./runs/{args.enc_name}/checkpoints/Res15-STEP-{str(args.enc_step)}.pt')
elif args.model == 'Quartz':
encoder = nemo_asr.JasperEncoder(**jasper_params["JasperEncoder"])
fc = LinearLayer(64 * 256) # TODO find shape from jasper_params
encoder.restore_from(f'./runs/{args.enc_name}/checkpoints/JasperEncoder-STEP-{str(args.enc_step)}.pt')
fc.restore_from(f'./runs/{args.enc_name}/checkpoints/LinearLayer-STEP-{str(args.enc_step)}.pt')
fc.freeze()
encoder.freeze()
l2_regularizer = L2Regularizer()
decoder = ClassificationNet(num_classes, args.hidden_size).to('cuda')
N = len(train_data_layer)
steps_per_epoch = math.ceil(N / float(batch_size) + 1)
logging.info("Steps per epoch : {0}".format(steps_per_epoch))
logging.info('Have {0} examples to train on.'.format(N))
ce_loss = nemo_asr.CrossEntropyLossNM()
logging.info('================================')
logging.info(f"Number of parameters in encoder: {encoder.num_weights}")
logging.info(f"Number of parameters in decoder: {decoder.num_weights}")
logging.info(
f"Total number of parameters in model: " f"{decoder.num_weights + encoder.num_weights}"
)
logging.info('================================')
"""BUILDING TRAIN GRAPH"""
audio_signal, audio_signal_len, commands, command_len = train_data_layer()
processed_signal, processed_signal_len = data_preprocessor(input_signal=audio_signal, length=audio_signal_len)
encoded, encoded_len = encoder(audio_signal=processed_signal, length=processed_signal_len)
if args.model == 'Quartz':
encoded = fc(embeddings=encoded)
encoded = l2_regularizer(embeds=encoded)
decoded = decoder(embeddings=encoded)
train_loss = ce_loss(logits=decoded, labels=commands)
"""BUILDING TEST GRAPH"""
test_audio_signal, test_audio_signal_len, test_commands, test_command_len = eval_data_layer()
test_processed_signal, test_processed_signal_len = data_preprocessor(
input_signal=test_audio_signal, length=test_audio_signal_len
)
test_encoded, test_encoded_len = encoder(audio_signal=test_processed_signal, length=test_processed_signal_len)
if args.model == 'Quartz':
test_encoded = fc(embeddings=test_encoded)
test_encoded = l2_regularizer(embeds=test_encoded)
test_decoded = decoder(embeddings=test_encoded)
test_loss = ce_loss(logits=test_decoded, labels=test_commands)
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[train_loss, decoded, commands],
print_func=partial(monitor_classification_training_progress, eval_metric=None),
get_tb_values=lambda x: [("loss", x[0])],
tb_writer=neural_factory.tb_writer,
)
tagname = 'TestSet'
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[test_loss, test_decoded, test_commands],
user_iter_callback=partial(process_classification_evaluation_batch, top_k=1),
user_epochs_done_callback=partial(process_classification_evaluation_epoch, eval_metric=1, tag=tagname),
eval_step=500,
tb_writer=neural_factory.tb_writer,
eval_at_start=False
)
chpt_callback = nemo.core.CheckpointCallback(
folder=neural_factory.checkpoint_dir,
step_freq=500,
checkpoints_to_keep=100
)
callbacks = [train_callback, eval_callback, chpt_callback]
lr_policy = CosineAnnealing(
total_steps=num_epochs * steps_per_epoch,
warmup_ratio=0.05,
min_lr=1e-4,
)
logging.info(f"Using `{lr_policy}` Learning Rate Scheduler")
neural_factory.train(
tensors_to_optimize=[train_loss],
callbacks=callbacks,
lr_policy=lr_policy,
optimizer="novograd",
optimization_params={
"num_epochs": num_epochs,
"max_steps": None,
"lr": lr,
"momentum": 0.95,
"betas": (0.98, 0.5),
"weight_decay": weight_decay,
"grad_norm_clip": None,
},
batches_per_step=1,
)