-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_model.py
177 lines (132 loc) · 5.87 KB
/
train_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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# File : train_model.py
# Author : Yan <yanwong@126.com>
# Date : 08.04.2020
# Last Modified Date: 27.04.2020
# Last Modified By : Yan <yanwong@126.com>
import argparse
import collections
import logging
import time
import os
import tensorflow as tf
import numpy as np
import model
import losses
import metrics
import data_utils
import configuration
logging.basicConfig(level=logging.INFO)
parser = argparse.ArgumentParser(description='Train the LSTM-CRF tagger.')
parser.add_argument('train_pattern',
help='File pattern of sharded TFRecord files containing '
'tf.Example protos for training.')
parser.add_argument('dev_pattern',
help='File pattern of sharded TFRecord files containing '
'tf.Example protos for validation.')
parser.add_argument('w2v',
help='The pre-trained word vector.')
parser.add_argument('vocab',
help='The vocabulary file containing all words')
parser.add_argument('ckpt_dir',
help='Directory for saving and loading checkpoints.')
# parser.add_argument('model_dir', default='lstm_crf_tagger',
# help='Directory for saving and loading model.')
# parser.add_argument('model_version', default='0001',
# help='Version of saved model.')
args = parser.parse_args()
model_config = configuration.ModelConfig()
train_config = configuration.TrainingConfig()
train_dataset = data_utils.create_dataset(args.train_pattern,
train_config.batch_size)
dev_dataset = data_utils.create_dataset(args.dev_pattern,
train_config.batch_size)
vocab = data_utils.load_vocab(args.vocab)
w2v = data_utils.load_w2v(args.w2v, vocab)
embeddings = data_utils.load_vocab_embeddings(w2v, vocab, model_config.d_word)
tagger = model.Model(embeddings, len(vocab), model_config)
optimizer = tf.keras.optimizers.SGD(lr=train_config.learning_rate,
clipvalue=train_config.clip_gradients)
train_loss = tf.keras.metrics.Mean(name='train_loss')
train_accuracy = tf.keras.metrics.Accuracy(name='train_accuracy')
train_metric = metrics.TaggerMetric(model_config.n_tags)
dev_accuracy = tf.keras.metrics.Accuracy(name='dev_accuracy')
dev_metric = metrics.TaggerMetric(model_config.n_tags)
ckpt = tf.train.Checkpoint(tagger=tagger, optimizer=optimizer)
ckpt_manager = tf.train.CheckpointManager(ckpt, args.ckpt_dir, max_to_keep=50)
# If you want to deploy model, put SavedModel here.
# model_path = os.path.join(args.model_dir, args.model_version)
if ckpt_manager.latest_checkpoint:
ckpt.restore(ckpt_manager.latest_checkpoint)
print('Latest checkpoint restored!')
def classification_report(metric):
metric_res = metric.result().numpy()
# print('Classification report:\n')
print('\taccuracy: ', metric_res[0])
print('\tprecison: ', metric_res[1])
print('\trecall: ', metric_res[2])
print('\tF1 score: ', metric_res[3])
def train_step(inp, tar):
# inp.shape == (batch_size, max_seq_len)
# tar.shape == (batch_size, max_seq_len)
padding_mask = data_utils.create_padding_mask(inp)
with tf.GradientTape() as tape:
pred, potentials = tagger(inp, True, padding_mask) # (batch_size, max_seq_len)
loss = losses.loss_function(tar, potentials, padding_mask,
tagger.crf_layer.trans_params)
gradients = tape.gradient(loss, tagger.trainable_variables)
optimizer.apply_gradients(zip(gradients, tagger.trainable_variables))
train_loss(loss)
train_accuracy(tar, pred, padding_mask)
train_metric(tar, pred, padding_mask)
def eval_step(inp, tar):
# inp.shape == (batch_size, max_seq_len)
# tar.shape == (batch_size, max_seq_len)
padding_mask = data_utils.create_padding_mask(inp)
pred, potentials = tagger(inp, False, padding_mask) # (batch_size, max_seq_len)
loss = losses.loss_function(tar, potentials, padding_mask,
tagger.crf_layer.trans_params)
dev_accuracy(tar, pred, padding_mask)
dev_metric(tar, pred, padding_mask)
def evaluate(dataset):
dev_accuracy.reset_states()
dev_metric.reset_states()
for batch, (inp, tar) in enumerate(dataset):
eval_step(inp, tar)
classification_report(dev_metric)
best_dev = -np.inf
for epoch in range(train_config.n_epochs):
start = time.time()
train_loss.reset_states()
train_accuracy.reset_states()
train_metric.reset_states()
dev_accuracy.reset_states()
dev_metric.reset_states()
for batch, (inp, tar) in enumerate(train_dataset):
train_step(inp, tar)
if batch % 50 == 0:
print('Epoch {} Batch {} Loss {:.4f} Accuracy {:.4f}'.format(
epoch + 1, batch, train_loss.result(), train_accuracy.result()))
classification_report(train_metric)
if batch % train_config.freq_eval == 0:
evaluate(dev_dataset)
macro_f1 = tf.math.reduce_mean(dev_metric.result()[3])
print('Eval accuracy {:.4f} Macro F1 {:.4f}'.format(
dev_accuracy.result(), macro_f1))
classification_report(dev_metric)
if macro_f1 > best_dev:
best_dev = macro_f1
print('New best dev F1 {:.4f}'.format(best_dev))
ckpt_save_path = ckpt_manager.save()
print('Saving checkpoint for epoch {} at {}.'.format(
epoch + 1, ckpt_save_path))
# Save model for deploying.
# Remember to use tagger.call() instead of tagger() in above inferences.
# tf.saved_model(tagger, model_path, signatures={'call': tagger.call})
# print('Saved model for epoch {} at {}.').format(epoch + 1, model_path)
print('Until now, best dev F1 {:.4f}'.format(best_dev))
print('Epoch {} Loss {:.4f} Accuracy {:.4f}'.format(
epoch + 1, train_loss.result(), train_accuracy.result()))
classification_report(train_metric)
print('Best dev F1 score {:.4f}'.format(best_dev))