forked from lucastheis/ride
/
train.py
282 lines (235 loc) · 10.2 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
"""
Train recurrent image density estimator on images.
"""
import os
import sys
import caffe
sys.path.append('./code')
from argparse import ArgumentParser
from copy import deepcopy
from numpy import sqrt, vstack, hstack, std, min, asarray
from numpy.random import permutation, rand, randn
from scipy.io import loadmat
from cmt.utils import random_select
from cmt.tools import imread
from ride import RIDE, PatchRIDE, MultiscaleRIDE, ColorRIDE
from tools import Experiment
def main(argv):
parser = ArgumentParser(argv[0], description=__doc__)
parser.add_argument('--patch_size', '-p', type=int, default=[8, 10, 12, 14, 16, 18, 20, 22], nargs='+')
parser.add_argument('--row_multiplier', '-R', type=int, default=[1], nargs='+',
help='Can be used to train on elongated patches.')
parser.add_argument('--col_multiplier', '-C', type=int, default=[1], nargs='+',
help='Can be used to train on elongated patches.')
parser.add_argument('--num_patches', '-P', type=int, default=None,
help='If given, subsample training data.')
parser.add_argument('--num_valid', '-V', type=int, default=0,
help='Number of training images used for validation error based early stopping.')
parser.add_argument('--finetune', '-F', type=int, default=[1], nargs='+',
help='Indicate iterations in which to finetune MCGSM with L-BFGS.')
parser.add_argument('--learning_rate', '-l', type=float, default=[1., .5, .1, .05, .01, 0.005, 0.001, 0.0005], nargs='+')
parser.add_argument('--momentum', '-m', type=float, default=[.9], nargs='+')
parser.add_argument('--batch_size', '-B', type=int, default=[50], nargs='+')
parser.add_argument('--nb_size', '-b', type=int, default=5,
help='Size of the causal neighborhood of pixels.')
parser.add_argument('--num_hiddens', '-n', type=int, default=64)
parser.add_argument('--num_components', '-c', type=int, default=32)
parser.add_argument('--num_scales', '-s', type=int, default=4)
parser.add_argument('--num_features', '-f', type=int, default=32)
parser.add_argument('--num_epochs', '-e', type=int, default=[1], nargs='+')
parser.add_argument('--precondition', '-Q', type=int, default=1)
parser.add_argument('--method', '-M', type=str, default=['SGD'], nargs='+')
parser.add_argument('--data', '-d', type=str, default='data/deadleaves_train.mat')
parser.add_argument('--noise', '-N', type=float, default=None,
help='Standard deviation of Gaussian noise added to data before training (as fraction of data standard deviation).')
parser.add_argument('--model', '-I', type=str, default='',
help='Start with this model as initialization. Other flags will be ignored.')
parser.add_argument('--add_layer', '-a', type=int, default=[0], nargs='+')
parser.add_argument('--train_top_layer', '-T', type=int, default=[0], nargs='+')
parser.add_argument('--train_means', '-S', type=int, default=[0], nargs='+')
parser.add_argument('--mode', '-q', type=str, default='CPU', choices=['CPU', 'GPU'])
parser.add_argument('--device', '-D', type=int, default=0)
parser.add_argument('--augment', '-A', type=int, default=1,
help='Increase training set size by transforming data.')
parser.add_argument('--overlap', '-O', type=int, default=[1], nargs='+')
parser.add_argument('--output', '-o', type=str, default='results/deadleaves/')
parser.add_argument('--patch_model', type=int, default=0,
help='Train a patch-based model instead of a stochastic process.')
parser.add_argument('--extended', '-X', type=int, default=0,
help='Use extended version of spatial LSTM.')
parser.add_argument('--multiscale', '-Y', type=int, default=0,
help='Apply recurrent image model to multiscale representation of images.')
parser.add_argument('--color', '-Z', type=int, default=0,
help='Use separate models to model color and grayscale values.')
args = parser.parse_args(argv[1:])
experiment = Experiment()
if args.mode.upper() == 'GPU':
caffe.set_mode_gpu()
caffe.set_device(args.device)
else:
caffe.set_mode_cpu()
# load data
if args.data.lower()[-4:] in ['.gif', '.png', '.jpg', 'jpeg']:
data = imread(args.data)[None]
data += rand(*data.shape)
else:
data = loadmat(args.data)['data']
if args.augment > 0:
data = vstack([data, data[:, :, ::-1]])
if args.augment > 1:
data = vstack([data, data[:, ::-1, :]])
if args.noise is not None:
# add noise as a means for regularization
data += randn(*data.shape) * (std(data, ddof=1) / args.noise)
if args.num_valid > 0:
if args.num_valid >= data.shape[0]:
print 'Cannot use {0} for validation, there are only {1} training images.'.format(
args.num_valid, data.shape[0])
return 1
# select subset for validation
idx = random_select(args.num_valid, data.shape[0])
data_valid = data[idx]
data = asarray([image for i, image in enumerate(data) if i not in idx])
print '{0} training images'.format(data.shape[0])
print '{0} validation images'.format(data_valid.shape[0])
patches_valid = []
patch_size = min([64, data.shape[1], data.shape[2]])
for i in range(0, data_valid.shape[1] - patch_size + 1, patch_size):
for j in range(0, data_valid.shape[2] - patch_size + 1, patch_size):
patches_valid.append(data_valid[:, i:i + patch_size, j:j + patch_size])
patches_valid = vstack(patches_valid)
if args.model:
# load pretrained model
results = Experiment(args.model)
model = results['model']
loss = [results['loss']]
if args.patch_model and not isinstance(model, PatchRIDE):
model = PatchRIDE(
model=model,
num_rows=args.patch_size[0],
num_cols=args.patch_size[0])
else:
# create recurrent image model
if args.patch_model:
model = PatchRIDE(
num_rows=args.patch_size[0],
num_cols=args.patch_size[0],
num_channels=data.shape[-1] if data.ndim > 3 else 1,
num_hiddens=args.num_hiddens,
nb_size=args.nb_size,
num_components=args.num_components,
num_scales=args.num_scales,
num_features=args.num_features,
model_class=ColorRIDE if args.color else RIDE)
if args.extended:
print 'Extended patch model not supported.'
return 0
if args.multiscale:
print 'Multiscale patch model not supported.'
return 0
else:
if args.multiscale:
if data.ndim > 3 and data.shape[-1] > 1:
print 'Multiscale color model not supported.'
return 0
model = MultiscaleRIDE(
num_hiddens=args.num_hiddens,
nb_size=args.nb_size,
num_components=args.num_components,
num_scales=args.num_scales,
num_features=args.num_features,
extended=args.extended > 0)
elif args.color:
if data.ndim < 4 or data.shape[-1] != 3:
print 'These images don\'t look like RGB images.'
return 0
model = ColorRIDE(
num_hiddens=args.num_hiddens,
nb_size=args.nb_size,
num_components=args.num_components,
num_scales=args.num_scales,
num_features=args.num_features,
extended=args.extended > 0)
else:
model = RIDE(
num_channels=data.shape[-1] if data.ndim > 3 else 1,
num_hiddens=args.num_hiddens,
nb_size=args.nb_size,
num_components=args.num_components,
num_scales=args.num_scales,
num_features=args.num_features,
extended=args.extended > 0)
loss = []
# compute initial performance
loss_valid = []
if args.num_valid > 0:
print 'Computing validation loss...'
loss_valid.append(model.evaluate(patches_valid))
model_copy = deepcopy(model)
for k, patch_size in enumerate(args.patch_size):
if args.multiscale:
patch_size *= 2
if k < len(args.add_layer):
for _ in range(args.add_layer[k]):
# add spatial LSTM to the network
model.add_layer()
# extract patches of given patch size
patches = []
row_size = patch_size * args.row_multiplier[k % len(args.row_multiplier)]
col_size = patch_size * args.col_multiplier[k % len(args.col_multiplier)]
if isinstance(model, PatchRIDE):
model.num_rows = row_size
model.num_cols = col_size
for i in range(0, data.shape[1] - row_size + 1, row_size / args.overlap[k % len(args.overlap)]):
for j in range(0, data.shape[2] - col_size + 1, col_size / args.overlap[k % len(args.overlap)]):
patches.append(data[:, i:i + row_size, j:j + col_size])
patches = vstack(patches)
# randomize order of patches
if args.num_patches is not None and args.num_patches < len(patches):
patches = patches[random_select(args.num_patches, len(patches))]
else:
patches = patches[permutation(len(patches))]
# determine batch size
if args.method[k % len(args.method)].upper() == 'SFO':
num_batches = int(max([25, sqrt(patches.shape[0]) / 5.]))
batch_size = patches.shape[0] // num_batches
else:
batch_size = args.batch_size[k % len(args.batch_size)]
if batch_size < 1:
raise RuntimeError('Too little data.')
print 'Patch size: {0}x{1}'.format(row_size, col_size)
print 'Number of patches: {0}'.format(patches.shape[0])
print 'Batch size: {0}'.format(batch_size)
# train recurrent image model
print 'Training...'
loss.append(
model.train(patches,
batch_size=batch_size,
method=args.method[k % len(args.method)],
num_epochs=args.num_epochs[k % len(args.num_epochs)],
learning_rate=args.learning_rate[k % len(args.learning_rate)],
momentum=args.momentum[k % len(args.momentum)],
precondition=args.precondition > 0,
train_top_layer=args.train_top_layer[k % len(args.train_top_layer)] > 0,
train_means=args.train_means[k % len(args.train_means)] > 0))
if args.finetune[k % len(args.finetune)]:
print 'Finetuning...'
model.finetune(patches, num_samples_train=1000000, max_iter=500)
if args.num_valid > 0:
print 'Computing validation loss...'
loss_valid.append(model.evaluate(patches_valid))
if loss_valid[-1] > loss_valid[-2]:
print 'Performance got worse. Stopping optimization.'
model = model_copy
break
print 'Copying model...'
model_copy = deepcopy(model)
experiment['batch_size'] = batch_size
experiment['args'] = args
experiment['model'] = model
experiment['loss_valid'] = loss_valid
experiment['loss'] = hstack(loss) if len(loss) > 0 else []
experiment.save(os.path.join(args.output, 'rim.{0}.{1}.xpck'))
return 0
if __name__ == '__main__':
sys.exit(main(sys.argv))