-
Notifications
You must be signed in to change notification settings - Fork 37
/
Copy pathtrain.py
213 lines (183 loc) · 8.99 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
# -----------------------------------------------------------------------
# train.py
# Trainer for a binarized CNN
#
# Creation Date : 04/Aug./2017
# Copyright (C) <2017> Hiroki Nakahara, All rights reserved.
#
# Released under the GPL v2.0 License.
#
# Acknowledgements:
# This source code is based on following projects:
#
# Chainer binarized neural network by Daisuke Okanohara
# https://github.com/hillbig/binary_net
# Various CNN models including Deep Residual Networks (ResNet)
# for CIFAR10 with Chainer by mitmul
# https://github.com/mitmul/chainer-cifar10
# -----------------------------------------------------------------------
import argparse
#import cPickle as pickle # python 2.7
import _pickle as pickle # python 3.5
import numpy as np
import os
import chainer
from chainer import optimizers
from chainer import serializers
import net2 # it will be generated by the GUINNESS GUI
import trainer
import time
import weight_clip
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='CIFAR-10 dataset trainer')
parser.add_argument('--gpu', '-g', type=int, default=-1,
help='GPU device ID (negative value indicates CPU)')
parser.add_argument('--model', '-m', type=str, default='bincnn', choices=['bincnn'],
help='Model name')
parser.add_argument('--batch_size', '-b', type=int, default=20,
help='Mini batch size')
parser.add_argument('--dataset', '-d', type=str, default='image.pkl',
help='Dataset image pkl file path')
parser.add_argument('--label', '-l', type=str, default='label.pkl',
help='Dataset label pkl file path')
parser.add_argument('--prefix', '-p', type=str, default='temp', # should be project name
help='Prefix of model parameter files')
parser.add_argument('--iter', type=int, default=10,
help='Training iteration')
parser.add_argument('--save_iter', type=int, default=0,
help='Iteration interval to save model parameter file.')
parser.add_argument('--lr_decay_iter', type=int, default=100,
help='Iteration interval to decay learning rate')
parser.add_argument('--weight_decay', type=float, default=0.0001,
help='Weight decay')
parser.add_argument('--optimizer', type=str, default='sgd', choices=['sgd', 'adam', 'momentum', 'delta'],
help='Optimizer name')
parser.add_argument('--lr', type=float, default=0.01,
help='Initial learning rate for SGD')
parser.add_argument('--alpha', type=float, default=0.00005,
help='Initial alpha for Adam')
parser.add_argument('--res_depth', type=int, default=18,
help='Depth of Residual Network')
parser.add_argument('--skip_depth', action='store_true',
help='Use stochastic depth in Residual Network')
parser.add_argument('--swapout', action='store_true',
help='Use swapout')
parser.add_argument('--seed', type=int, default=1,
help='Random seed')
parser.add_argument('--dim', type=int, default=3,
help='Dimension (default RGB, that is, 3)')
parser.add_argument('--siz', type=int, default=32,
help='ImageSiz (default 32, that is, 32x32)')
parser.add_argument('--guinness', type=str, default='./hoge', # should be project name
help='Prefix of model parameter files for the GUINNESS flow')
parser.add_argument('--resume', type=str, default='no',
help='Resume traning, if pre-trained model exists')
args = parser.parse_args()
np.random.seed(args.seed)
log_file_path = '{}_log.csv'.format(args.prefix)
# lr_decay_iter = map(int, args.lr_decay_iter.split(','))
if args.prefix is None:
model_prefix = '{}_{}'.format(args.model, args.optimizer)
else:
model_prefix = args.prefix
# load image dataset
print('loading dataset %s' % args.dataset)
with open(args.dataset, 'rb') as f:
images = pickle.load(f)
index = np.random.permutation(len(images['train']))
threshold = np.int32(len(images['train'])/10*9)
train_index = index[:threshold]
valid_index = index[threshold:]
train_x = images['train'][train_index].astype(np.float32)
valid_x = images['train'][valid_index].astype(np.float32)
test_x = images['test'].astype(np.float32)
print("[INFO] #TRAIN DATA: %7d" % len(train_x))
print("[INFO] #VALID DATA: %7d" % len(valid_x))
print("[INFO] #TEST DATA: %7d" % len(test_x))
# load label dataset
with open(args.label, 'rb') as f:
labels = pickle.load(f)
train_y = labels['train'][train_index].astype(np.int32)
valid_y = labels['train'][valid_index].astype(np.int32)
test_y = labels['test'].astype(np.int32)
# generate testbench (test_img.txt) for C/C++ code
idx = 0
image = test_x
# extract only one image
image1 = image[idx]
# generate text file as a bench marck
bench_img = image1.transpose(1,2,0)
bench_img = bench_img.reshape(-1,)
fname = 'test_img.txt'
print(' Test Image Fileout -> %s' % fname)
np.savetxt(fname, bench_img,fmt="%.0f",delimiter=",")
# start training
print('start training')
cifar_net = net2.CNN() # modified
# resume pre-trained model, if exist
if args.resume == 'yes':
print(" Resume Pre-Trained Model")
serializers.load_npz('{}.model'.format(model_prefix), cifar_net)
if args.optimizer == 'sgd':
print("optimizer: SGD")
optimizer = optimizers.SGD(lr=args.lr)
elif args.optimizer == 'momentum':
print("optimizer: momentum SGD")
optimizer = optimizers.MomentumSGD(lr=args.lr)
elif args.optimizer == 'delta':
print("optimizer: AdaDelta")
optimizer = optimizers.AdaDelta()
else:
print("optimizer: Adam")
optimizer = optimizers.Adam(alpha=args.alpha)
optimizer.setup(cifar_net)
if args.weight_decay > 0:
optimizer.add_hook(chainer.optimizer.WeightDecay(args.weight_decay))
optimizer.add_hook(weight_clip.WeightClip())
cifar_trainer = trainer.CifarTrainer(cifar_net, optimizer, args.iter, args.batch_size, args.gpu)
state = {'best_valid_error': 100, 'best_test_error': 100, 'clock': time.clock()}
def on_epoch_done(epoch, n, o, loss, acc, valid_loss, valid_acc, test_loss, test_acc):
error = 100 * (1 - acc)
valid_error = 100 * (1 - valid_acc)
test_error = 100 * (1 - test_acc)
print('epoch {} done'.format(epoch))
print('train loss: {} error: {}'.format(loss, error))
print('valid loss: {} error: {}'.format(valid_loss, valid_error))
print('test loss: {} error: {}'.format(test_loss, test_error))
if valid_error < state['best_valid_error']:
serializers.save_npz('{}.model'.format(model_prefix), n)
serializers.save_npz('{}.state'.format(model_prefix), o)
state['best_valid_error'] = valid_error
state['best_test_error'] = test_error
if args.save_iter > 0 and (epoch + 1) % args.save_iter == 0:
serializers.save_npz('{}_{}.model'.format(model_prefix, epoch + 1), n)
serializers.save_npz('{}_{}.state'.format(model_prefix, epoch + 1), o)
# prevent divergence when using identity mapping model
if args.model == 'identity_mapping' and epoch < 9:
o.lr = 0.01 + 0.01 * (epoch + 1)
# if len(lr_decay_iter) == 1 and (epoch + 1) % lr_decay_iter[0] == 0 or epoch + 1 in lr_decay_iter:
# Note, "lr_decay_iter" should be a list object to store a training schedule,
# However, to keep up with the Python3.5, I changed to an integer value...
if (epoch + 1) % args.lr_decay_iter == 0 and epoch > 1:
if hasattr(optimizer, 'alpha'):
o.alpha *= 0.1
else:
o.lr *= 0.1
clock = time.clock()
print('elapsed time: {}'.format(clock - state['clock']))
state['clock'] = clock
with open(log_file_path, 'a') as f:
f.write('{},{},{},{},{},{},{}\n'.format(epoch + 1, loss, error, valid_loss, valid_error, test_loss, test_error))
if args.resume == 'no':
print(" Create %s as a New Logfile" % log_file_path)
with open(log_file_path, 'w') as f:
f.write('epoch,train loss,train acc,valid loss,valid acc,test loss,test acc\n')
else:
print(" Overwrite Existing Logfile %s" % log_file_path)
cifar_trainer.fit(train_x, train_y, valid_x, valid_y, args.siz, args.dim, test_x, test_y, on_epoch_done)
print('best test error: {}'.format(state['best_test_error']))
with open("train_status.txt", 'w') as f:
f.write("stop")
# -----------------------------------------------------------------------
# END OF PROGRAM
# -----------------------------------------------------------------------