-
Notifications
You must be signed in to change notification settings - Fork 27
/
Copy pathtrain.py
190 lines (162 loc) · 9.9 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
from __future__ import print_function
import argparse
import os
import random
import numpy as np
from keras.callbacks import TensorBoard
from keras.optimizers import SGD, Adam, Adagrad
from tqdm import tqdm
from model import DIIN
from optimizers.l2optimizer import L2Optimizer
from util import ChunkDataManager
class Gym(object):
def __init__(self,
model,
train_data,
test_data,
dev_data,
optimizers,
logger,
models_save_dir):
self.model = model
self.logger = logger
''' Data '''
self.train_data = train_data
self.test_data = test_data
self.dev_data = dev_data
self.model_save_dir = models_save_dir
if not os.path.exists(self.model_save_dir):
os.mkdir(self.model_save_dir)
''' Optimizers '''
self.optimizers = optimizers
self.optimizer_id = -1
self.current_optimizer = None
self.current_switch_step = -1
def switch_optimizer(self):
self.optimizer_id += 1
if self.optimizer_id >= len(self.optimizers):
print('Finished training...')
exit(0)
self.current_optimizer, self.current_switch_step = self.optimizers[self.optimizer_id]
self.model.compile(optimizer=self.current_optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
self.logger.set_model(self.model)
print('Switching to number {} optimizer'.format(self.current_optimizer))
def train(self, batch_size=70, eval_interval=500, shuffle=True):
print('train:\t', [d.shape for d in self.train_data])
print('test:\t', [d.shape for d in self.test_data])
print('dev:\t', [d.shape for d in self.dev_data])
# Initialize optimizer
self.switch_optimizer()
self.model.summary()
# Start training
train_step, eval_step, no_progress_steps = 0, 0, 0
train_batch_start = 0
best_loss = 1000.
while True:
if shuffle:
random.shuffle(list(zip(train_data)))
train_inputs = train_data[:-1]
train_labels = train_data[-1]
# Evaluate
test_loss, dev_loss = self.evaluate(eval_step=eval_step, batch_size=batch_size)
eval_step += 1
# Switch optimizer if it's necessary
no_progress_steps += 1
if dev_loss < best_loss:
best_loss = dev_loss
no_progress_steps = 0
if no_progress_steps >= self.current_switch_step:
self.switch_optimizer()
no_progress_steps = 0
# Train eval_interval times
for _ in tqdm(range(eval_interval)):
[loss, acc] = model.train_on_batch(
[train_input[train_batch_start: train_batch_start + batch_size] for train_input in train_inputs],
train_labels[train_batch_start: train_batch_start + batch_size])
self.logger.on_epoch_end(epoch=train_step, logs={'train_acc': acc, 'train_loss': loss})
train_step += 1
train_batch_start += batch_size
if train_batch_start > len(train_inputs[0]):
train_batch_start = 0
# Shuffle the data after the epoch ends
if shuffle:
random.shuffle(list(zip(train_data)))
def evaluate(self, eval_step, batch_size=None):
[test_loss, test_acc] = model.evaluate(self.test_data[:-1], self.test_data[-1], batch_size=batch_size)
[dev_loss, dev_acc] = model.evaluate(self.dev_data[:-1], self.dev_data[-1], batch_size=batch_size)
self.logger.on_epoch_end(epoch=eval_step, logs={'test_acc': test_acc, 'test_loss': test_loss})
self.logger.on_epoch_end(epoch=eval_step, logs={'dev_acc': dev_acc, 'dev_loss': dev_loss})
model.save(self.model_save_dir + 'epoch={}-tloss={}-tacc={}.model'.format(eval_step, test_loss, test_acc))
return test_loss, dev_loss
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', default=70, help='Batch size', type=int)
parser.add_argument('--eval_interval', default=500, help='Evaluation Interval (#batches)', type=int)
parser.add_argument('--char_embed_size', default=8, help='Size of character embedding', type=int)
parser.add_argument('--char_conv_filters', default=100, help='Number of character conv filters', type=int)
parser.add_argument('--char_conv_kernel', default=5, help='Size of char convolution kernel', type=int)
parser.add_argument('--dropout_initial_keep_rate', default=1., help='Initial keep rate of decaying dropout', type=float)
parser.add_argument('--dropout_decay_rate', default=0.977, help='Decay rate of dropout', type=float)
parser.add_argument('--dropout_decay_interval', default=10000, help='Dropout decay interval', type=int)
parser.add_argument('--l2_full_step', default=100000, help='Number of steps for full L2 penalty', type=float)
parser.add_argument('--l2_full_ratio', default=9e-5, help='L2 full penalty', type=float)
parser.add_argument('--l2_diference_penalty', default=1e-3, help='L2 penalty applied on weight difference', type=float)
parser.add_argument('--first_scale_down_ratio', default=0.3, help='First scale down ratio (DenseNet)', type=float)
parser.add_argument('--transition_scale_down_ratio', default=0.5, help='Transition scale down ratio (DenseNet)', type=float)
parser.add_argument('--growth_rate', default=20, help='Growth rate (DenseNet)', type=int)
parser.add_argument('--layers_per_dense_block', default=8, help='Layers in one Dense block (DenseNet)', type=int)
parser.add_argument('--dense_blocks', default=3, help='Number of Dense blocks (DenseNet)', type=int)
parser.add_argument('--labels', default=3, help='Number of output labels', type=int)
parser.add_argument('--load_dir', default='data', help='Directory of the data', type=str)
parser.add_argument('--models_dir', default='models/', help='Where to save models', type=str)
parser.add_argument('--logdir', default='logs', help='Tensorboard logs dir', type=str)
parser.add_argument('--word_vec_path', default='data/word-vectors.npy', help='Save path word vectors', type=str)
parser.add_argument('--omit_word_vectors', action='store_true')
parser.add_argument('--omit_chars', action='store_true')
parser.add_argument('--omit_syntactical_features', action='store_true')
parser.add_argument('--omit_exact_match', action='store_true')
parser.add_argument('--train_word_embeddings', action='store_true')
args = parser.parse_args()
''' Prepare data '''
word_embedding_weights = np.load(args.word_vec_path)
train_data = ChunkDataManager(load_data_path=os.path.join(args.load_dir, 'train')).load()
test_data = ChunkDataManager(load_data_path=os.path.join(args.load_dir, 'test')).load()
dev_data = ChunkDataManager(load_data_path=os.path.join(args.load_dir, 'dev')).load()
''' Getting dimensions of the input '''
chars_per_word = train_data[3].shape[-1] if not args.omit_chars else 0
syntactical_feature_size = train_data[5].shape[-1] if not args.omit_syntactical_features else 0
''' Prepare the model and optimizers '''
adam = L2Optimizer(Adam(), args.l2_full_step, args.l2_full_ratio, args.l2_diference_penalty)
adagrad = L2Optimizer(Adagrad(), args.l2_full_step, args.l2_full_ratio, args.l2_diference_penalty)
sgd = L2Optimizer(SGD(lr=3e-3), args.l2_full_step, args.l2_full_ratio, args.l2_diference_penalty)
model = DIIN(p=train_data[0].shape[-1], # or None
h=train_data[1].shape[-1], # or None
include_word_vectors=not args.omit_word_vectors,
word_embedding_weights=word_embedding_weights,
train_word_embeddings=args.train_word_embeddings,
include_chars=not args.omit_chars,
chars_per_word=chars_per_word,
char_embedding_size=args.char_embed_size,
char_conv_filters=args.char_conv_filters,
char_conv_kernel_size=args.char_conv_kernel,
include_syntactical_features=not args.omit_syntactical_features,
syntactical_feature_size=syntactical_feature_size,
include_exact_match=not args.omit_exact_match,
dropout_initial_keep_rate=args.dropout_initial_keep_rate,
dropout_decay_rate=args.dropout_decay_rate,
dropout_decay_interval=args.dropout_decay_interval,
first_scale_down_ratio=args.first_scale_down_ratio,
transition_scale_down_ratio=args.transition_scale_down_ratio,
growth_rate=args.growth_rate,
layers_per_dense_block=args.layers_per_dense_block,
nb_dense_blocks=args.dense_blocks,
nb_labels=args.labels)
''' Initialize Gym for training '''
gym = Gym(model=model,
train_data=train_data, test_data=test_data, dev_data=dev_data,
optimizers=[(adam, 3), (adagrad, 4), (sgd, 15)],
logger=TensorBoard(log_dir=args.logdir),
models_save_dir=args.models_dir)
gym.train(batch_size=args.batch_size, eval_interval=args.eval_interval, shuffle=True)