-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
305 lines (234 loc) · 11.1 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
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
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
import torch
import torch.nn as nn
import os, glob
import optuna
import numpy as np
from CNN import CNN
from DataHelper import BertHelper
from torch.utils.data import DataLoader
from tqdm import tqdm
from collections import OrderedDict
from sklearn.metrics import f1_score
from transformers import AdamW, get_linear_schedule_with_warmup
gpu = torch.device('cuda:0')
# compute precision@k function
def precision_k(target_b, outputs_b, k=1):
p_k_batch = np.empty(0, dtype=np.float32)
for target, outputs in zip(target_b, outputs_b):
target = np.where(target > 0)[0]
outputs = np.argpartition(-outputs, k)[:k]
accuracy = sum([(i in target) for i in outputs])
p_k_value = [accuracy != 0.0 and accuracy / k or 0.0]
p_k_batch = np.concatenate([p_k_batch, np.array(p_k_value)])
return p_k_batch
class Model() :
def __init__(self, params) :
self.params = params
torch.manual_seed(params['seed'])
if params['cuda']:
torch.cuda.manual_seed(params['seed'])
# making dataloader, model and etc...
def build(self) :
# Dataloaderの定義
train_helper = BertHelper(self.params['train_path'])
self.train_loader = DataLoader(
train_helper,
batch_size = self.params['batch_size'],
shuffle = True
)
val_helper = BertHelper(self.params['val_path'])
self.val_loader = DataLoader(
val_helper,
batch_size = self.params['batch_size'],
shuffle = True
)
test_helper = BertHelper(self.params['test_path'])
self.test_loader = DataLoader(
test_helper,
batch_size = self.params['batch_size'],
shuffle = True
)
print('train sample:' + str(len(train_helper)) + ', val sample:' + str(len(val_helper)) + ', test sample:' + str(len(test_helper)))
print('finished data regularization!')
print('create model...', end = '')
# CNNモデルの定義
self.model = CNN(self.params)
if self.params['cuda'] :
self.model.cuda()
# 損失関数の定義
self.criterion = nn.BCEWithLogitsLoss()
# Optimizerの定義,及び重み減衰を適応させるパラメータの選択
no_decay = ['bias', 'LayerNorm.weight']
optimizer_parameters = [
{'params' : [p for i, p in self.model.named_parameters() if not any(j in i for j in no_decay)], 'weight_decay' : self.params['weight_decay']},
{'params' : [p for i, p in self.model.named_parameters() if any(j in i for j in no_decay)], 'weight_decay' : 0.0},
]
self.optimizer = AdamW(optimizer_parameters, lr = self.params['learning_rate'])
# Schedulerの定義
num_training_steps = len(self.train_loader) * self.params['epoch']
num_warmup_steps = len(self.train_loader) * self.params['warmup_steps']
self.scheduler = get_linear_schedule_with_warmup(self.optimizer, num_warmup_steps, num_training_steps)
print('done!')
print('training start')
# Parameter search function
def tuning(self, trial) :
# CNNにおけるフィルタ数
cnn_filter_num = trial.suggest_int('cnn_filter_num', 1, 4)
# CNNにおけるフィルタのサイズ, filter_size ** filter_num
cnn_filter_size = trial.suggest_int('cnn_filter_sizes', 2, 4)
cnn_filter_sizes = [cnn_filter_size ** i for i in range(1, cnn_filter_num + 1)]
# CNNにおけるConvolution層のストライド幅
cnn_conv_stride = trial.suggest_int('cnn_conv_stride', 1, 8)
# CNNにおけるMaxPooling層のストライド幅
cnn_pool_stride = trial.suggest_int('cnn_pool_stride', 1, 8)
# CNNにおけるDropout
#cnn_dropout1 = trial.suggest_categorical('cnn_dropout1', [False, 0.25, 0.5, 0.75])
cnn_dropout1 = False
#cnn_dropout2 = trial.suggest_categorical('cnn_dropout2', [False, 0.25, 0.5, 0.75])
cnn_dropout2 = False
# CNNにおけるチャンネル数
cnn_out_channels = trial.suggest_categorical('cnn_out_channels', [2 ** i for i in range(1, 8)])
# CNNにおける全結合層の隠れ次元
cnn_hidden_dim1 = trial.suggest_categorical('cnn_hidden_dim1', [2 ** i for i in range(5, 11)])
# 学習率
learning_rate = trial.suggest_loguniform('learning_rate', 0.0000001, 0.1)
# WarmupSchedulerのWarmup地点
#warmup_steps = trial.suggest_categorical('warmup_steps', [0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
warmup_steps = 0
# Epoch
#self.params['epoch'] = trial.suggest_categorical('epoch', [1, 2, 3, 4, 5, 6, 7, 8, 9, 10])
# 閾値
#threshold = trial.suggest_categorical('threshold', [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0])
threshold = 0.5
# 重み減衰
weight_decay = trial.suggest_loguniform('weight_decay', 0.00001, 0.1)
learning_params = {
'batch_size' : 2,
'learning_rate' : learning_rate,
'warmup_steps' : warmup_steps,
'threshold' : threshold,
'weight_decay' : weight_decay
}
self.params.update(learning_params)
cnn_params = {
'cnn_out_channels' : cnn_out_channels,
'cnn_filter_sizes' : cnn_filter_sizes,
'cnn_hidden_dim1' : cnn_hidden_dim1,
'cnn_conv_stride' : cnn_conv_stride,
'cnn_pool_stride' : cnn_pool_stride,
'cnn_dropout' : [cnn_dropout1, cnn_dropout2],
}
self.params.update(cnn_params)
self.build()
score = 0
for epoch in range(1, self.params['epoch'] + 1) :
self.train(epoch)
if epoch % self.params['epoch'] == 0 :
score, val_loss, result = self.test('val')
return 1.0 - score
# normal train&test function
def run(self) :
self.build()
max_score = 0
best_epoch = 0
scores = []
bad_counter = 0
for epoch in range(1, self.params['epoch'] + 1) :
self.train(epoch)
# early stopping
if epoch % 1 == 0 :
score, val_loss, result = self.test('val')
scores.append(score)
torch.save(self.model.state_dict(), '{}.pkl'.format(epoch))
if scores[-1] > max_score :
max_score = scores[-1]
best_epoch = epoch
bad_counter = 0
else:
bad_counter += 1
if bad_counter == self.params['patience'] :
break
files = glob.glob('*.pkl')
for file in files:
epoch_nb = int(file.split('.')[0])
if epoch_nb < best_epoch :
os.remove(file)
files = glob.glob('*.pkl')
for file in files:
epoch_nb = int(file.split('.')[0])
if epoch_nb > best_epoch :
os.remove(file)
print("Optimization Finished!")
print('Loading {}th epoch'.format(best_epoch))
self.model.load_state_dict(torch.load('{}.pkl'.format(best_epoch)))
score, loss_mean, result = self.test('test')
def train(self, epoch) :
with tqdm(self.train_loader) as bar :
for i, (texts, labels) in enumerate(bar) :
if self.params['cuda'] :
texts = texts.to(gpu)
labels = labels.to(gpu)
self.model.train()
self.optimizer.zero_grad()
output = self.model(texts)
train_loss = self.criterion(output, labels)
train_loss.backward()
self.optimizer.step()
self.scheduler.step()
bar.set_description('Train epoch %d' % epoch)
bar.set_postfix(OrderedDict(loss = train_loss.item()))
def test(self, mode) :
self.model.eval()
y_pre = np.empty((0, 103), dtype = np.float32)
y_true = np.empty((0, 103), dtype = np.float32)
p_1 = np.empty(0, dtype = np.float32)
p_3 = np.empty(0, dtype = np.float32)
p_5 = np.empty(0, dtype = np.float32)
if mode == 'val' :
loader = self.val_loader
else :
loader = self.test_loader
with tqdm(loader) as bar :
result = {}
epoch_loss = []
for i, (texts, labels) in enumerate(bar) :
if self.params['cuda'] :
texts = texts.to(gpu)
labels = labels.to(gpu)
with torch.no_grad() :
output = self.model(texts)
test_loss = self.criterion(output, labels)
epoch_loss.append(test_loss.item())
output = torch.sigmoid(output)
output = output.cpu().detach_().numpy().copy()
labels = labels.cpu().detach_().numpy().copy()
output_threshold = (output > self.params['threshold'])
y_pre = np.append(y_pre, output_threshold, axis = 0)
y_true = np.append(y_true, labels, axis = 0)
p_1_batch = precision_k(labels, output, k=1)
p_3_batch = precision_k(labels, output, k=3)
p_5_batch = precision_k(labels, output, k=5)
bar.set_description(mode)
bar.set_postfix(OrderedDict(loss = test_loss.item()))
p_1 = np.concatenate([p_1, p_1_batch])
p_3 = np.concatenate([p_3, p_3_batch])
p_5 = np.concatenate([p_5, p_5_batch])
loss_mean = np.mean(epoch_loss)
result['loss'] = loss_mean
print('loss/val : ' + str(loss_mean))
p_1 = np.mean(p_1)
p_3 = np.mean(p_3)
p_5 = np.mean(p_5)
print('p@1 : ' + str(p_1))
print('p@3 : ' + str(p_3))
print('p@5 : ' + str(p_5))
result['p@1'] = p_1
result['p@3'] = p_3
result['p@5'] = p_5
macroF1 = f1_score(y_true, y_pre, average = 'macro', zero_division = 0)
microF1 = f1_score(y_true, y_pre, average = 'micro', zero_division = 0)
print('macro f1 : ' + str(macroF1))
print('micro f1 : ' + str(microF1))
result['macro'] = macroF1
result['micro'] = microF1
return macroF1, loss_mean, result