-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_pawsx.py
203 lines (184 loc) · 8.47 KB
/
train_pawsx.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
import os
os.environ['TF_KERAS'] = '1'
# os.environ["CUDA_VISIBLE_DEVICES"] = '0'
import tensorflow as tf
import numpy as np
import random
from tensorflow import keras as K
from tensorflow.keras.layers import Lambda
from transformers import XLMRobertaTokenizer
from tqdm import tqdm
from bert4keras.models import build_transformer_model
from bert4keras.layers import Loss
from tensorflow.keras.callbacks import ModelCheckpoint
my_seed = 1234
np.random.seed(my_seed)
random.seed(my_seed)
tf.random.set_seed(my_seed)
os.environ['PYTHONHASHSEED'] = str(my_seed)
SEQ_LEN = 128
BATCH_SIZE = 32
EPOCHS = 8
LR = 5e-5
print(LR, BATCH_SIZE)
pretrained_path = 'PCT/xlm_roberta_base/'
config_path = os.path.join(pretrained_path, 'xlm_roberta_base_config.json')
checkpoint_path = os.path.join(pretrained_path, 'xlm_roberta_base.ckpt')
# vocab_path = os.path.join(pretrained_path, 'vocab.txt')
#
tpu = tf.distribute.cluster_resolver.TPUClusterResolver() # TPU detection
print('Running on TPU ', tpu.cluster_spec().as_dict()['worker'])
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
strategy = tf.distribute.experimental.TPUStrategy(tpu)
# strategy = tf.distribute.MirroredStrategy()
print("REPLICAS: ", strategy.num_replicas_in_sync)
tokenizer = XLMRobertaTokenizer.from_pretrained('xlm-roberta-base', pad_token='<s>')
tokenizer.padding_side = 'right'
special_tokens = ['<unused{}>'.format(i) for i in range(200)]
tokenizer.add_special_tokens({'additional_special_tokens': special_tokens})
label_map = {'1': 63736, '0': 157}
template = {'en': 'Question:{} paraphrase or not? Answer:<mask>', 'zh': '问题：{} 是否转述? 答案：<mask>',
'fr': 'Question:{} paraphrase ou pas? Réponse:<mask>',
'es': 'Pregunta: {} ¿parafrasear o no? Respuesta: <mask>',
'de': 'Frage:{} paraphrasieren oder nicht? Antwort:<mask>', 'ja': '質問：{}言い換えるかどうか?回答：<mask>',
'ko': '질문:{} 의역 여부는? 답:<mask>'
}
language_list = ['fr', 'es', 'de', 'ja', 'ko', 'zh']
test_data = {}
def load_data(path, is_training=False, type=None):
global tokenizer
idx = 0
indices, labels, indices_target, labels_target = [], [], [], []
with open(path, 'r') as reader:
for line in tqdm(reader.readlines()):
if not line: continue
if line.startswith('premise'): continue
if line.startswith('id\t'): continue
array = line.strip().split('\t')
label = array[3]
premise = array[1]
hypothesis = template['en'].format(array[2])
label_id = label_map[label]
token_ids = tokenizer.encode(premise, hypothesis, max_length=SEQ_LEN,
truncation='longest_first', padding='max_length')
assert len(token_ids) == SEQ_LEN
if 250001 not in token_ids:
template_tmp = template['en'].split('{}')[1]
template_ids = tokenizer.encode(template_tmp)
length = len(template_ids)
token_ids = token_ids[:-length + 1]
token_ids.extend(template_ids[1:])
target = np.equal(np.array(token_ids), 250001).astype('int')
assert np.sum(target) == 1
target = target * label_id
# print(target)
indices.append(token_ids)
labels.append(target)
if is_training:
flag = random.randint(0, len(language_list) - 1)
language_sample = language_list[flag]
template_cur = template[language_sample]
hypothesis_ = template_cur.format(array[2])
token_ids_ = tokenizer.encode(premise, hypothesis_, max_length=SEQ_LEN,
truncation='longest_first', padding='max_length')
if 250001 not in token_ids_:
template_tmp = template_cur.split('{}')[1]
template_ids = tokenizer.encode(template_tmp)
length = len(template_ids)
token_ids_ = token_ids_[:-length + 1]
token_ids_.extend(template_ids[1:])
target_ = np.equal(np.array(token_ids_), 250001).astype('int')
assert np.sum(target_) == 1
target_ = target_ * label_id
assert len(token_ids_) == SEQ_LEN
indices_target.append(token_ids_)
labels_target.append(target_)
if len(array) == 5 and type == 'test':
language = array[-1]
if language in test_data:
test_data[language]['token_ids'].append(token_ids)
test_data[language]['label'].append(label_id)
else:
test_data[language] = {'token_ids': [token_ids], 'label': [label_id]}
if idx < 5:
print("*** Example ***")
print("guid: %s" % (idx))
print("tokens: %s" % " ".join(
[x for x in tokenizer.decode(token_ids)]))
print("input_ids: %s" % " ".join([str(x) for x in token_ids]))
print("label: %s (id = %s)" % (label, str(labels[idx])))
idx += 1
# if len(indices) > 1000:
# break
indices = np.array(indices)
labels = np.array(labels)
if not is_training:
return [indices, indices, labels, labels], None
else:
indices_target = np.array(indices_target)
labels_target = np.array(labels_target)
return [indices, indices_target, labels, labels_target], None
train_path = 'PCT/datasets/x-final/en/train.tsv'
valid_path = 'PCT/datasets/x-final/valid_all.txt'
test_path = 'PCT/datasets/x-final/test_all.txt'
#
train_x, train_y = load_data(train_path, is_training=True)
valid_x, valid_y = load_data(valid_path)
# test_x, test_y = load_data(test_path, type='test')
with strategy.scope():
class CrossEntropy(Loss):
def compute_loss(self, inputs, mask=None):
y_true, y_pred = inputs
y_mask = K.backend.cast(K.backend.not_equal(y_true, 0), K.backend.floatx())
accuracy = K.metrics.sparse_categorical_accuracy(y_true, y_pred)
accuracy = K.backend.sum(accuracy * y_mask) / (K.backend.sum(y_mask) + 1e-6)
self.add_metric(accuracy, name='accuracy', aggregation='mean')
loss = K.backend.sparse_categorical_crossentropy(y_true, y_pred)
loss = K.backend.sum(loss * y_mask) / (K.backend.sum(y_mask) + 1e-6)
return loss
bert = build_transformer_model(
config_path=config_path,
checkpoint_path=checkpoint_path,
model='xlm_roberta',
return_keras_model=True,
sequence_length=SEQ_LEN,
segment_vocab_size=0,
with_mlm=True
)
y_in = K.layers.Input(shape=(None,))
y_in_target = K.layers.Input(shape=(None,))
y_label = K.layers.Input(shape=(None,))
y_label_target = K.layers.Input(shape=(None,))
context_emb = bert(y_in)
context_emb_target = bert(y_in_target)
inputs = [y_in, y_in_target, y_label, y_label_target]
outputs = CrossEntropy(1)([y_in, context_emb])
outputs_ = CrossEntropy(1)([y_in_target, context_emb_target])
outputs_predict = Lambda(
lambda x: K.backend.sum(x[0] * K.backend.expand_dims(K.backend.cast(K.backend.equal(x[1], 250001),
'float'), -1), axis=1))(
[context_emb, inputs[0]])
outputs_predict_target = Lambda(
lambda x: K.backend.sum(x[0] * K.backend.expand_dims(K.backend.cast(K.backend.equal(x[1], 250001),
'float'), -1), axis=1))(
[context_emb_target, inputs[1]])
optimizer = K.optimizers.RMSprop(LR)
model = K.models.Model(inputs, [outputs, outputs_])
model_predict = K.models.Model(inputs[0], outputs_predict)
model.add_loss(Lambda(lambda x: K.backend.mean(K.losses.kld(x[0], x[1]) + K.losses.kld(x[1], x[0])))(
[outputs_predict, outputs_predict_target]))
model.compile(
optimizer,
)
model.summary()
checkpoint = ModelCheckpoint(filepath='PCT/model.hdf5', monitor='val_accuracy', verbose=True,
save_best_only=True, save_weights_only=True, mode='max')
model.fit(
train_x, train_y,
epochs=EPOCHS,
batch_size=BATCH_SIZE,
validation_data=(valid_x, valid_y),
callbacks=[checkpoint]
)
# model.save_weights('PCT/model.hdf5')