This repository has been archived by the owner on Jul 25, 2018. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
executable file
·335 lines (314 loc) · 12.2 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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import argparse
import collections
import datetime
import logging
import os
import os.path as osp
import pickle
import sys
import matplotlib
matplotlib.use('Agg')
from chainer import cuda
from chainer import optimizers as O
from chainer import serializers
from chainer import Variable
import numpy as np
from skimage.io import imread
from skimage.transform import resize
from apc_od import draw_loss_curve
from apc_od import get_raw
from apc_od import im_preprocess
from apc_od import im_to_blob
from apc_od import mask_to_roi
from apc_od import raw_to_mask_path
from apc_od import roi_preprocess
from apc_od import tile_ae_encoded
from apc_od import tile_ae_inout
here = osp.dirname(osp.abspath(__file__))
class Trainer(object):
def __init__(
self,
optimizers,
model,
model_name,
is_supervised,
crop_roi,
batch_size,
log_dir,
log_file,
on_gpu,
):
self.optimizers = optimizers
self.model = model
self.model_name = model_name
self.is_supervised = is_supervised
self.crop_roi = crop_roi
self.batch_size = batch_size
self.log_dir = log_dir
self.log_file = log_file
self.on_gpu = on_gpu
def batch_loop(self, x_data, t_data, train):
N = len(x_data)
# train loop
sum_loss = collections.defaultdict(float)
sum_accuracy = 0 if self.is_supervised else None
perm = np.random.permutation(N)
for i in range(0, N, self.batch_size):
print('train.Trainer.batch_loop: i: {}'.format(i))
x_batch = x_data[perm[i:i + self.batch_size]]
t_batch = t_data[perm[i:i + self.batch_size]]
if self.on_gpu:
x_batch = cuda.to_gpu(x_batch)
t_batch = cuda.to_gpu(t_batch)
volatile = 'off' if train else 'on'
x = Variable(x_batch, volatile=volatile)
t = Variable(t_batch, volatile=volatile)
for j in xrange(len(self.optimizers)):
self.optimizers[j].zero_grads()
if self.is_supervised:
inputs = [x, t]
else:
inputs = [x]
self.model.train = train
if train:
losses = self.model(*inputs)
if losses is None:
continue
if not isinstance(losses, collections.Sequence):
losses = [losses]
self.model.to_gpu()
for j, loss in enumerate(losses):
loss.backward()
self.optimizers[j].update()
else:
losses = self.model(*inputs)
if not isinstance(losses, collections.Sequence):
losses = [losses]
for j, loss in enumerate(losses):
sum_loss[j] += len(t.data) * float(loss.data)
if self.is_supervised:
sum_accuracy += len(t.data) * float(self.model.accuracy.data)
y_data = self.model.y.data
if self.on_gpu:
x_batch = cuda.to_cpu(x_batch)
y_data = cuda.to_cpu(y_data)
return sum_loss, sum_accuracy, x_batch, y_data
@staticmethod
def dataset_to_xt_data(dataset, crop_roi):
"""Convert dataset to x_data and t_data"""
x_data = []
for raw_path in dataset.filenames:
raw = im_preprocess(imread(raw_path))
if crop_roi:
mask_path = raw_to_mask_path(raw_path)
roi = mask_to_roi(im_preprocess(imread(mask_path)))
raw = raw[roi[0]:roi[2], roi[1]:roi[3]]
raw = resize(raw, (128, 128), preserve_range=True)
x_data.append(im_to_blob(raw))
x_data = np.array(x_data, dtype=np.float32)
t_data = dataset.target.astype(np.int32)
return x_data, t_data
def main_loop(self, n_epoch=10, save_interval=None, save_encoded=True):
save_interval = save_interval or (n_epoch // 10)
train_data = get_raw(which_set='train')
test_data = get_raw(which_set='test')
N_train = len(train_data.filenames)
N_test = len(test_data.filenames)
logging.info('converting dataset to x and t data')
train_x, train_t = self.dataset_to_xt_data(train_data, self.crop_roi)
test_x, test_t = self.dataset_to_xt_data(test_data, self.crop_roi)
for epoch in xrange(0, n_epoch):
# train
sum_loss, sum_accuracy, _, _ = \
self.batch_loop(train_x, train_t, train=True)
for loss_id, sl in sorted(sum_loss.items()):
mean_loss = sl / N_train
msg = 'epoch:{:02d}; train mean loss{}={};'\
.format(epoch, loss_id, mean_loss)
if self.is_supervised:
mean_accuracy = sum_accuracy / N_train
msg += ' accuracy={};'.format(mean_accuracy)
logging.info(msg)
print(msg)
# test
sum_loss, sum_accuracy, x_batch, y_batch = \
self.batch_loop(test_x, test_t, train=False)
for loss_id, sl in sorted(sum_loss.items()):
mean_loss = sl / N_test
msg = 'epoch:{:02d}; test mean loss{}={};'\
.format(epoch, loss_id, mean_loss)
if self.is_supervised:
mean_accuracy = sum_accuracy / N_test
msg += ' accuracy={};'.format(mean_accuracy)
logging.info(msg)
print(msg)
# save model and input/encoded/decoded
if epoch % save_interval == (save_interval - 1):
print('epoch:{:02d}; saving'.format(epoch))
# save model
model_path = osp.join(
self.log_dir,
'{name}_model_{epoch}.h5'.format(
name=self.model_name, epoch=epoch))
serializers.save_hdf5(model_path, self.model)
# save optimizer
for i, opt in enumerate(self.optimizers):
opt_path = osp.join(
self.log_dir,
'{name}_optimizer_{epoch}_{i}.h5'.format(
name=self.model_name, epoch=epoch, i=i))
serializers.save_hdf5(opt_path, opt)
# save x_data
x_path = osp.join(self.log_dir, 'x_{}.pkl'.format(epoch))
with open(x_path, 'wb') as f:
pickle.dump(x_batch, f) # save x
if not self.is_supervised:
x_hat_path = osp.join(self.log_dir,
'x_hat_{}.pkl'.format(epoch))
with open(x_hat_path, 'wb') as f:
pickle.dump(y_batch, f) # save x_hat
tile_ae_inout(
x_batch, y_batch,
osp.join(self.log_dir, 'X_{}.jpg'.format(epoch)))
if save_encoded:
x = Variable(cuda.to_gpu(x_batch), volatile=True)
z = self.model.encode(x)
tile_ae_encoded(
cuda.to_cpu(z.data),
osp.join(self.log_dir,
'x_encoded_{}.jpg'.format(epoch)))
for i in xrange(len(self.optimizers)):
draw_loss_curve(
loss_id=i,
logfile=self.log_file,
outfile=osp.join(self.log_dir, 'loss_curve{}.jpg'.format(i)),
no_acc=not self.is_supervised,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
'supervised_or_not', type=str, choices=['supervised', 'unsupervised'],
help='do supervised or unsupervised training')
parser.add_argument('--epoch', type=int, default=50,
help='number of recursion (default: 50)')
parser.add_argument('--no-logging', action='store_true',
help='logging to tmp dir')
parser.add_argument('--save-interval', type=int, default=None,
help='save interval of x and x_hat')
parser.add_argument('-m', '--model', required=True, help='name of model')
args = parser.parse_args()
n_epoch = args.epoch
save_interval = args.save_interval
is_supervised = True if args.supervised_or_not == 'supervised' else False
on_gpu = True
is_pipeline = False
batch_size = 10
save_encoded = False
crop_roi = False
optimizers = [O.Adam()]
if is_supervised:
if args.model == 'VGG_mini_ABN':
from apc_od.models import VGG_mini_ABN
model = VGG_mini_ABN()
if on_gpu:
model.to_gpu()
optimizers[0].setup(model)
crop_roi = True
elif args.model == 'CAEOnesRoiVGG':
from apc_od.pipeline import CAEOnesRoiVGG
is_pipeline = True
batch_size = 10
initial_roi = np.array([0, 0, 356, 534])
logging.info('initial_roi: {}'.format(initial_roi))
initial_roi = roi_preprocess(initial_roi)
# setup model
model = CAEOnesRoiVGG(initial_roi=initial_roi, learning_rate=0.2, learning_n_sample=300)
if on_gpu:
model.to_gpu()
optimizers = [O.Adam(), O.Adam()]
optimizers[0].setup(model.cae_ones1)
optimizers[1].setup(model.vgg2)
# load trained models
serializers.load_hdf5(os.path.join(here, 'cae_ones_model.h5'),
model.cae_ones1)
serializers.load_hdf5(os.path.join(here, 'vgg_model.h5'),
model.vgg2)
# load optimizers state
serializers.load_hdf5(os.path.join(here, 'cae_ones_optimizer.h5'),
optimizers[0])
serializers.load_hdf5(os.path.join(here, 'vgg_optimizer.h5'),
optimizers[1])
else:
sys.stderr.write('Unsupported model: {}\n'.format(args.model))
sys.exit(1)
else:
# unsupervised
if args.model == 'CAE':
from apc_od.models import CAE
save_encoded = True
model = CAE()
if on_gpu:
model.to_gpu()
optimizers[0].setup(model)
elif args.model == 'CAEOnes':
from apc_od.models import CAEOnes
model = CAEOnes()
if on_gpu:
model.to_gpu()
optimizers[0].setup(model)
elif args.model == 'CAEPool':
from apc_od.models import CAEPool
save_encoded = True
model = CAEPool()
if on_gpu:
model.to_gpu()
optimizers[0].setup(model)
elif args.model == 'StackedCAE':
from apc_od.models import StackedCAE
save_encoded = True
model = StackedCAE()
if on_gpu:
model.to_gpu()
optimizers[0].setup(model)
else:
sys.stderr.write('Unsupported model: {}\n'.format(args.model))
sys.exit(1)
timestamp = datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S')
# setup for logging
if args.no_logging:
import tempfile
log_dir = tempfile.mkdtemp()
else:
log_dir = osp.join(here, '../logs/{}_{}'.format(timestamp, args.model))
log_dir = osp.realpath(osp.abspath(log_dir))
os.mkdir(log_dir)
log_file = osp.join(log_dir, 'log.txt')
logging.basicConfig(
format='%(asctime)s [%(levelname)s] %(message)s',
filename=log_file,
level=logging.DEBUG,
)
logging.info('args: {};'.format(args))
msg = 'logging in {};'.format(log_dir)
logging.info(msg)
print(msg)
trainer = Trainer(
optimizers=optimizers,
model=model,
model_name=args.model,
is_supervised=is_supervised,
crop_roi=crop_roi,
batch_size=batch_size,
log_dir=log_dir,
log_file=log_file,
on_gpu=on_gpu,
)
trainer.main_loop(
n_epoch=n_epoch,
save_interval=save_interval,
save_encoded=save_encoded,
)
if __name__ == '__main__':
main()