-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
196 lines (152 loc) · 7.32 KB
/
train.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 pandas as pd
import numpy as np
import yaml
from data_helper import *
from sent_simtf import SentenceSimilarity
import tensorflow as tf
import json
import os
import itertools
import shutil
from tensorflow import set_random_seed
# from dpcnn import DPCNN
set_random_seed(2018)
np.random.seed(2018)
def main():
with open('./config.yml', encoding='utf-8') as f:
config = yaml.load(f)
learning_rate = config['model_params']['learning_rate']
max_len = config['max_len']
# num_classes = config['model_params']['num_classes']
batch_size = config['model_params']['batch_size']
min_count = config['data_params']['min_count']
data_path = config['data_params']['data_path']
val_path = config['data_params']['val_path']
kernel_size = config['model_params']['kernel_size']
num_filters = config['model_params']['num_filters']
embedding_size = config['model_params']['embedding_size']
hidden_size = config['model_params']['hidden_size']
dropout_keep_prob = config['model_params']['keep_prob']
epochs = config['model_params']['epochs']
embedding_path = config['embedding_path']
gpu_id = config['gpu_id']
# if os.path.exists(config['ckpt_file']):
# shutil.rmtree(config['ckpt_file'])
# os.mkdir(config['ckpt_file'])
# else:
# os.mkdir(config['ckpt_file'])
if not os.path.exists(config['ckpt_file']):
os.mkdir(config['ckpt_file'])
# 加载数据
data, word2id = load_data(data_path, config['ckpt_file'], min_count=min_count)
# data = pd.read_csv(data_path,delimiter='\t',header=None)
vocab_size = len(word2id)
# val
# val = pd.read_csv(val_path,delimiter='\t',header=None)
val, val_word2id = load_data(val_path, config['ckpt_file'], min_count=5, char=True, write_vocab=False)
# 对句子进行编码
train_id = pd.Series(list(map(lambda x: string2id(x, word2id, char=True), data[1])))
data[2] = train_id
val_id = pd.Series(list(map(lambda x: string2id(x, word2id, char=True), val[1])))
val[2] = val_id
##
train_data = data
train_data = train_data.sample(frac=1)
x_train = np.array(list(train_data[2]))
y_train = list(train_data[0])
num_classes = len(set(y_train))
print('num_classes', num_classes)
####
os.environ['CUDA_VISIBLE_ORDER'] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_id
with tf.Session() as sess:
ss = SentenceSimilarity(max_len, embedding_size, vocab_size, hidden_size, num_classes, learning_rate)
# ss = DPCNN(vocab_size, kernel_size, num_filters, max_len, embedding_size,num_classes,learning_rate)
saver = tf.train.Saver()
if os.path.exists(config['ckpt_file'] + 'checkpoint'):
print("Restore Variables from checkpoint!")
saver.restore(sess, tf.train.latest_checkpoint(config['ckpt_file']))
else:
print("Initializer variables!")
sess.run(tf.global_variables_initializer())
accs = {'top1': [], 'top3': []}
if config['use_embedding']:
embedding = pretrained_embedding(embedding_path, word2id, embedding_size)
assign_Embedding = tf.assign(ss.Embedding, embedding)
sess.run(assign_Embedding)
highest, boundary = 0.0, 0.65
for epoch in range(epochs):
counter, loss, total_num, total_correct = 0.0, 0.0, 0.0, 0.0
for x_batch, y_batch, text_len in batch_buckets(x_train, y_train, batch_size, buckets_len=[15, 20, 40, 80]):
# print("x_batch",x_batch)
# print('y_batch',y_batch)
counter += 1
# print('counter',counter)
feed_dict = {ss.input_x: x_batch, ss.input_y: y_batch, ss.x_lens: text_len,
ss.dropout_keep_prob: dropout_keep_prob, ss.batch_size: len(x_batch)}
cur_loss, _, accuracy = sess.run([ss.loss, ss.train_op, ss.accuracy], feed_dict=feed_dict)
# cur_loss,_,accuracy,cos_value,idxs,cos = sess.run([ss.loss,ss.train_op,ss.accuracy,ss.y_true_pred,ss.idxs,ss.cosine],
# feed_dict=feed_dict)
# print('y_batch:',y_batch[0:5])
# print('cos shape:',cos.shape)
# print('idxs:',idxs.shape)
# print('cos_value:',cos_value[0:5])
# print('idxs_value:',idxs[0:5])
# print('cos:',cos[0:5])
loss += cur_loss * len(x_batch)
total_num += len(x_batch)
total_correct += accuracy * len(x_batch)
if counter % 100 == 0:
# print("cur_loss",cur_loss)
# print("accuracy",accuracy)
top1_val, top3_val, top1_threshold, top3_threshold = do_eval(sess, ss, val, batch_size, boundary)
print('Epoch %d/%d\tbatch %d\ttrain_loss:%.3f\ttrain_accuracy:%.3f' % (
epoch, epochs, counter, loss / total_num, total_correct / total_num))
accs['top1'].append(top1_val)
accs['top3'].append(top3_val)
if top1_val >= highest:
highest = top1_val
save_file = config['ckpt_file'] + 'model.ckpt'
saver.save(sess, save_file, global_step=epoch)
print('find new model,top1_val: %s, top3_val: %s, top1大于 %s: %s, top3 大于%s' % (
top1_val, top3_val, boundary, top1_threshold, top3_threshold))
json.dump({'accs': accs, 'highest_top1': highest}, open('valid_amsoftmax.log', 'w'), indent=4)
def do_eval(sess, model, val, batch_size, thre): # cosine [Batch,num_class],在这里num_class为batch,[Batch,Batch]
id2g = dict(zip(val.index - val.index[0], val[0]))
text_len = list(map(lambda x: len(x), val[2]))
maxlen = max(text_len)
# print(maxlen)
def func(sent, maxlen):
_ = sent[:maxlen] + [0] * (maxlen - len(sent))
return _
token_ids = list(map(lambda x: func(x, maxlen), val[2]))
# top1_num,top3_num = 0.0,0.0
feed_dict = {model.input_x: token_ids, model.x_lens: text_len, model.dropout_keep_prob: 1.0}
rnn_outputs = sess.run(model.outputs_rnn, feed_dict=feed_dict) # [Batch,hidden_size]
# print(rnn_outputs[0:5])
cosine = np.matmul(rnn_outputs, np.transpose(rnn_outputs)) # [Batch,Batch]
# print('cosine shape',cosine[0:5])
max_index = np.array(list(map(lambda x: np.argsort(x)[::-1][:4], cosine)))
top_cosine = np.array(list(map(lambda x: np.sort(x)[::-1][:4], cosine)))
new_result = np.vectorize(lambda x: id2g[x])(max_index) # 转换组别
threshold = np.full((max_index.shape[0],), thre)
# print(threshold)
# top = top_cosine[:,1]
_ = new_result[:, 0] != new_result[:, 0]
_t = new_result[:, 0] != new_result[:, 0]
for i in range(3):
mask = new_result[:, 0] == new_result[:, i + 1]
top = top_cosine[:, i + 1]
# print(top[0:10])
real_correct = (top * mask >= threshold)
_t += real_correct
_ = _ + mask
if i + 1 == 1:
top1_acc = 1. * _.sum() / len(_)
top1_threshold = _t.sum() / len(_t)
elif i + 1 == 3:
top3_acc = 1. * _.sum() / len(_)
top3_threshold = _t.sum() / len(_t)
return top1_acc, top3_acc, top1_threshold, top3_threshold
if __name__ == '__main__':
main()