-
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train-classifier.py
261 lines (221 loc) · 10.9 KB
/
train-classifier.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
import argparse
import json
import os
import numpy as np
import tensorflow.compat.v1 as tf
import time
import tqdm
from tensorflow.core.protobuf import rewriter_config_pb2
from model import modeling
from model.modeling import BertConfig, BertModel
from run_finetune import get_masked_lm_output,get_next_sentence_output
from encode_bpe import BPEEncoder_ja
parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, default='RoBERTa-ja_small', help='pretrained model directory.')
parser.add_argument('--input_dir', type=str, required=True, help='input texts.')
parser.add_argument('--train_by_line', action='store_true', help='split file by lines.')
parser.add_argument('--batch_size', type=int, default=1, help='training batch size.')
parser.add_argument('--run_name', type=str, default='RoBERTa-ja_classifier', help='save model name.')
parser.add_argument('--save_every', type=int, default=2, help='save every N epoch.')
parser.add_argument('--num_epochs', type=int, default=20, help='training epochs.')
parser.add_argument('--learning_rate', type=float, default=0.00001, help='learning rate.')
parser.add_argument('--gpu', default='0', help='visible gpu number.')
CHECKPOINT_DIR = 'checkpoint'
def maketree(path):
try:
os.makedirs(path)
except:
pass
def main():
args = parser.parse_args()
if os.path.isfile(args.model+'/hparams.json'):
with open(args.model+'/hparams.json') as f:
bert_config_params = json.load(f)
else:
raise ValueError('invalid model name.')
vocab_size = bert_config_params['vocab_size']
max_seq_length = bert_config_params['max_position_embeddings']
batch_size = args.batch_size
save_every = args.save_every
num_epochs = args.num_epochs
EOT_TOKEN = vocab_size - 4
MASK_TOKEN = vocab_size - 3
CLS_TOKEN = vocab_size - 2
SEP_TOKEN = vocab_size - 1
with open('ja-bpe.txt', encoding='utf-8') as f:
bpe = f.read().split('\n')
with open('emoji.json', encoding='utf-8') as f:
emoji = json.loads(f.read())
enc = BPEEncoder_ja(bpe, emoji)
keys = [f for f in os.listdir(args.input_dir) if os.path.isdir(args.input_dir+'/'+f)]
keys = sorted(keys)
num_labels = len(keys)
input_contexts = []
input_keys = []
idmapping_dict = {}
for i,f in enumerate(keys):
n = 0
for t in os.listdir(f'{args.input_dir}/{f}'):
if os.path.isfile(f'{args.input_dir}/{f}/{t}'):
with open(f'{args.input_dir}/{f}/{t}', encoding='utf-8') as fn:
if args.train_by_line:
for p in fn.readlines():
tokens = enc.encode(p.strip())[:max_seq_length-2]
tokens = [CLS_TOKEN]+tokens+[SEP_TOKEN]
if len(tokens) < max_seq_length:
tokens.extend([0]*(max_seq_length-len(tokens)))
input_contexts.append(tokens)
input_keys.append(i)
n += 1
else:
p = fn.read()
tokens = enc.encode(p.strip())[:max_seq_length-3]
tokens = [CLS_TOKEN]+tokens+[EOT_TOKEN,SEP_TOKEN]
if len(tokens) < max_seq_length:
tokens.extend([0]*(max_seq_length-len(tokens)))
input_contexts.append(tokens)
input_keys.append(i)
n += 1
print(f'{args.input_dir}/{f} mapped for id_{i}, read {n} contexts.')
idmapping_dict[f] = i
input_indexs = np.random.permutation(len(input_contexts))
bert_config = BertConfig(**bert_config_params)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = args.gpu
with tf.Session(config=config) as sess:
input_ids = tf.placeholder(tf.int32, [None, None])
input_mask = tf.placeholder(tf.int32, [None, None])
segment_ids = tf.placeholder(tf.int32, [None, None])
masked_lm_positions = tf.placeholder(tf.int32, [None, None])
masked_lm_ids = tf.placeholder(tf.int32, [None, None])
masked_lm_weights = tf.placeholder(tf.float32, [None, None])
next_sentence_labels = tf.placeholder(tf.int32, [None])
model = BertModel(
config=bert_config,
is_training=True,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=segment_ids,
use_one_hot_embeddings=False)
output = model.get_sequence_output()
(_,_,_) = get_masked_lm_output(
bert_config, model.get_sequence_output(), model.get_embedding_table(),
masked_lm_positions, masked_lm_ids, masked_lm_weights)
(_,_,_) = get_next_sentence_output(
bert_config, model.get_pooled_output(), next_sentence_labels)
saver = tf.train.Saver()
ckpt = tf.train.latest_checkpoint(args.model)
saver.restore(sess, ckpt)
train_vars = tf.trainable_variables()
restored_weights = {}
for i in range(len(train_vars)):
restored_weights[train_vars[i].name] = sess.run(train_vars[i])
labels = tf.placeholder(tf.int32, [None, ])
output_layer = model.get_pooled_output()
if int(tf.__version__[0]) > 1:
hidden_size = output_layer.shape[-1]
else:
hidden_size = output_layer.shape[-1].value
output_weights = tf.get_variable(
"output_weights", [num_labels, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
"output_bias", [num_labels], initializer=tf.zeros_initializer())
with tf.variable_scope("loss"):
output_layer = tf.nn.dropout(output_layer, keep_prob=0.9)
logits = tf.matmul(output_layer, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
probabilities = tf.nn.softmax(logits, axis=-1)
log_probs = tf.nn.log_softmax(logits, axis=-1)
one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32)
per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1)
loss = tf.reduce_mean(per_example_loss)
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
train_vars = tf.trainable_variables()
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summaries = tf.summary.scalar('loss', loss)
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
hparams_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'hparams.json')
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
with open(hparams_path, 'w') as fp:
fp.write(json.dumps(bert_config_params))
idmaps_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'idmaps.json')
with open(idmaps_path, 'w') as fp:
fp.write(json.dumps(idmapping_dict))
sess.run(tf.global_variables_initializer()) # init output_weights
restored = 0
for k,v in restored_weights.items():
for i in range(len(train_vars)):
if train_vars[i].name == k:
assign_op = train_vars[i].assign(v)
sess.run(assign_op)
restored += 1
assert restored == len(restored_weights), 'fail to restore model.'
saver = tf.train.Saver(var_list=tf.trainable_variables())
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(
sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
avg_loss = (0.0, 0.0)
start_time = time.time()
def sample_feature(i):
last = min((i+1)*batch_size,len(input_indexs))
_input_ids = [input_contexts[idx] for idx in input_indexs[i*batch_size:last]]
_input_masks = [[1]*len(input_contexts[idx])+[0]*(max_seq_length-len(input_contexts[idx])) for idx in input_indexs[i*batch_size:last]]
_segments = [[1]*len(input_contexts[idx])+[0]*(max_seq_length-len(input_contexts[idx])) for idx in input_indexs[i*batch_size:last]]
_labels = [input_keys[idx] for idx in input_indexs[i*batch_size:last]]
return {
input_ids:_input_ids,
input_mask:_input_masks,
segment_ids:_segments,
masked_lm_positions:np.zeros((len(_input_ids),0), dtype=np.int32),
masked_lm_ids:np.zeros((len(_input_ids),0), dtype=np.int32),
masked_lm_weights:np.ones((len(_input_ids),0), dtype=np.float32),
next_sentence_labels:np.zeros((len(_input_ids),), dtype=np.int32),
labels:_labels
}
try:
for ep in range(num_epochs):
if ep % args.save_every == 0:
save()
prog = tqdm.tqdm(range(0,len(input_contexts)//batch_size,1))
for i in prog:
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summaries),
feed_dict=sample_feature(i))
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
prog.set_description(
'[{ep} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(
ep=ep,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
counter += 1
except KeyboardInterrupt:
print('interrupted')
save()
save()
if __name__ == '__main__':
main()