/
train.py
376 lines (314 loc) · 14.3 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
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
import os
import argparse
import sys
import json
import math
import time
import numpy as np
import cv2
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
from sklearn.utils import shuffle
from sklearn.metrics import silhouette_score
import multiprocessing as mp
import matplotlib as mpl
mpl.use('Agg')
import matplotlib.pyplot as plt
from docrec.models.affinet import AffiNET
CPUS_RATIO = 0.5 # % of cpus used for dataset processing
SEED = 0 # <= change this in case of multiple runs
np.random.seed(SEED)
tf.set_random_seed(SEED)
def cohen_d(x, y):
nx = x.size
ny = y.size
dof = nx + ny - 2
return (x.mean() - y.mean()) / np.sqrt(((nx - 1) * x.std(ddof=1) ** 2 + (ny - 1 ) * y.std(ddof=1) ** 2) / dof)
class Dataset:
def __init__(self, args, mode='train', sess=None):
assert mode in ['train', 'val']
lines = open('{}/{}.txt'.format(args.samples_dir, mode)).readlines()
info = json.load(open('{}/info.json'.format(args.samples_dir), 'r'))
num_negatives = info['stats']['negatives_{}'.format(mode)]
num_positives = info['stats']['positives_{}'.format(mode)]
num_samples_per_class = min(num_positives, num_negatives)
self.num_samples = 2 * num_samples_per_class
self.curr_epoch = 1
self.num_epochs = args.num_epochs
self.curr_batch = 1
self.batch_size = args.batch_size
self.sample_size = info['params']['sample_size']
self.num_batches = math.ceil(self.num_samples / self.batch_size)
self.sess = sess
assert self.num_samples > self.batch_size
def _parse_function(filename, label):
''' Parse function. '''
image_string = tf.read_file(filename)
image = tf.image.decode_jpeg(image_string, channels=3, dct_method='INTEGER_ACCURATE') # works with png too
image = tf.image.convert_image_dtype(image, tf.float32)
image_left, image_right = tf.split(image, 2, axis=1) # split horizontally images into halves
# if mode == 'train':
# return image_left, image_right, tf.one_hot(label, NUM_CLASSES)
return image_left, image_right, label
# load samples' filenames and labels
count = {'0': 0, '1': 0} # balancing classes
labels = []
filenames = []
for line in lines:
filename, label = line.split()
if count[label] < num_samples_per_class:
filenames.append(filename)
labels.append(int(label))
count[label] += 1
# dataset iterator
dataset_tf = tf.data.Dataset.from_tensor_slices((filenames, labels))
if mode == 'train':
dataset_tf = dataset_tf.shuffle(len(filenames), seed=SEED) # important => reshuffle each iteraction
dataset_tf = dataset_tf.map(_parse_function, num_parallel_calls=max(1, int(CPUS_RATIO * mp.cpu_count())))
dataset_tf = dataset_tf.batch(args.batch_size).repeat(args.num_epochs)
self.next_batch_op = dataset_tf.make_one_shot_iterator().get_next()
def next_batch(self):
return self.sess.run(self.next_batch_op)
def train(args):
# reset the default graph and init a session
tf.reset_default_graph()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# create traindata directory if not existing (this directory is not syncronized in github)
os.makedirs('traindata', exist_ok=True)
# load samples dataset / samples size
print('loading training samples :: ', end='')
sys.stdout.flush()
dataset = Dataset(args, mode='train', sess=sess)
sample_height, sample_width = dataset.sample_size
print('num_samples={} sample_size={}x{}'.format(dataset.num_samples, sample_height, sample_width))
# placeholders
# 1) images (channels last)
images_left_ph = tf.placeholder(
tf.float32, name='left_ph', shape=(None, sample_height, sample_width // 2, 3)
)
images_right_ph = tf.placeholder(
tf.float32, name='right_ph', shape=(None, sample_height, sample_width - sample_width // 2, 3)
)
# 2) labels
# compatible 1 (+)
# uncompatible 0 (-)
labels_ph = tf.placeholder(tf.float32, name='labels_ph', shape=(None,))
# models
net_left = AffiNET(
images_left_ph, feat_dim=args.feat_dim, feat_layer=args.feat_layer, mode='train',
pretrained=False, base_arch='squeezenet', channels_first=False, activation=args.activation,
sample_height=sample_height, model_scope='left', seed=SEED, sess=sess
)
net_right = AffiNET(
images_right_ph, feat_dim=args.feat_dim, feat_layer=args.feat_layer, mode='train',
pretrained=False, base_arch='squeezenet', channels_first=False, activation=args.activation,
sample_height=sample_height, model_scope='right', seed=SEED, sess=sess
)
# features: batch x n_features dimensions, where n_features = height * dim_feat
pred_left = tf.squeeze(net_left.features, 1)
pred_right = tf.squeeze(net_right.features, 1)
# contrastive loss function
# http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf
# https://stackoverflow.com/questions/41172500/how-to-implement-metrics-learning-using-siamese-neural-network-in-tensorflow
d = tf.reduce_sum(tf.square(tf.subtract(pred_left, pred_right)), 1) # sum feature-wise
d_sqrt = tf.sqrt(d)
loss_op = (1.0 - labels_ph) * tf.square(tf.maximum(0.0, args.margin - d_sqrt)) + labels_ph * d
loss_op = 0.5 * tf.reduce_mean(loss_op) # averaging on batch
# optimizer (SGD)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=args.learning_rate)
train_op = optimizer.minimize(loss_op)
# grads_and_vars = optimizer.compute_gradients(loss_op)
# # train_op = optimizer.apply_gradients(grads_and_vars)
# init variables
sess.run(tf.global_variables_initializer())
# training loop
start = time.time()
losses_group = []
losses = []
steps = []
global_step = 0
num_steps_per_epoch = math.ceil(dataset.num_samples / args.batch_size)
total_steps = args.num_epochs * num_steps_per_epoch
for epoch in range(1, args.num_epochs + 1):
for step in range(1, num_steps_per_epoch + 1):
# batch data
images_left, images_right, labels = dataset.next_batch()
# train
loss, _, dists = sess.run(
[loss_op, train_op, d_sqrt],
feed_dict={images_left_ph: images_left, images_right_ph: images_right, labels_ph: labels}
)
losses_group.append(loss)
if (step % 100 == 0) or (step == num_steps_per_epoch):
losses.append(np.mean(losses_group))
steps.append(global_step)
elapsed = time.time() - start
remaining = elapsed * (total_steps - global_step) / global_step
print('[{:.2f}%] step={}/{} epoch={} loss={:.3f} :: {:.2f}/{:.2f} seconds lr={}'.format(
100 * global_step / total_steps, global_step, total_steps, epoch,
np.mean(losses_group), elapsed, remaining, args.learning_rate
))
losses_group = []
labels_arr = np.array(labels.tolist())
dists_arr = np.array(dists.tolist())
print('dist(-)={:.4f} dist(+)={:.4f}'.format(
dists_arr[labels_arr == 0].mean(), dists_arr[labels_arr == 1].mean())
)
# increment global step
global_step += 1
# save epoch model
net_left.save_weights('traindata/{}/model/left/{}.npy'.format(args.model_id, epoch))
net_right.save_weights('traindata/{}/model/right/{}.npy'.format(args.model_id, epoch))
plt.plot(steps, losses)
plt.savefig('traindata/{}/loss.png'.format(args.model_id))
sess.close()
def validate(args):
# reset the default graph and init a session
tf.reset_default_graph()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# load samples dataset / samples size
print('loading validation samples :: ', end='')
sys.stdout.flush()
dataset = Dataset(args, mode='val', sess=sess)
sample_height, sample_width = dataset.sample_size
print('num_samples={} sample_size={}x{}'.format(dataset.num_samples, sample_height, sample_width))
# placeholders
# 1) images
images_left_ph = tf.placeholder(
tf.float32, name='left_ph', shape=(None, sample_height, sample_width // 2, 3)
) # channels last
images_right_ph = tf.placeholder(
tf.float32, name='right_ph', shape=(None, sample_height, sample_width - sample_width // 2, 3)
) # channels last
# models
net_left = AffiNET(
images_left_ph, feat_dim=args.feat_dim, feat_layer=args.feat_layer, mode='val',
base_arch='squeezenet', channels_first=False, activation=args.activation,
sample_height=sample_height, model_scope='left', sess=sess
)
net_right = AffiNET(
images_right_ph, feat_dim=args.feat_dim, feat_layer=args.feat_layer, mode='val',
base_arch='squeezenet', channels_first=False, activation=args.activation,
sample_height=sample_height, model_scope='right', sess=sess
)
# features: batch x n_features dimensions
pred_left = tf.squeeze(net_left.features, 1) # batch x (height * dim_feat)
pred_right = tf.squeeze(net_right.features, 1) # batch x (height * dim_feat)
# distances
dists_op = tf.reduce_sum(tf.square(tf.subtract(pred_left, pred_right)), 1) # sum feature-wise
# init variables
sess.run(tf.global_variables_initializer())
# validation loop
max_accuracy = -1
best_epoch = -1
accuracies = []
num_steps_per_epoch = math.ceil(dataset.num_samples / args.batch_size)
total_steps = args.num_epochs * num_steps_per_epoch
for epoch in range(1, args.num_epochs + 1):
net_left.load_weights('traindata/{}/model/left/{}.npy'.format(args.model_id, epoch))
net_right.load_weights('traindata/{}/model/right/{}.npy'.format(args.model_id, epoch))
dists = []
labels = []
for step in range(1, num_steps_per_epoch + 1):
images_left, images_right, labels_step = dataset.next_batch()
dists_step = sess.run(
dists_op, feed_dict={images_left_ph: images_left, images_right_ph: images_right}
)
dists.append(dists_step)
labels.append(labels_step)
# calculate accuracy
dists_arr = np.concatenate(dists)
labels_arr = np.concatenate(labels)
accuracy = cohen_d(dists_arr[labels_arr == 0], dists_arr[labels_arr == 1])
accuracies.append(accuracy)
if accuracy > max_accuracy: # grab the highest / best epoch
max_accuracy = accuracy
best_epoch = epoch
print('[{:.2f}%] epoch={}/{} (best={}) accuracy={:.5f} (max={:.5f})'.format(
100 * epoch / args.num_epochs, epoch, args.num_epochs, best_epoch,
accuracy, max_accuracy
))
plt.cla()
plt.plot(np.arange(1, args.num_epochs + 1), accuracies)
# plt.xticks(np.arange(1, args.num_epochs + 1))
plt.vlines(best_epoch, 0.0, 1.0, colors='r', linestyles='dashed')
plt.savefig('traindata/{}/accuracy.png'.format(args.model_id))
sess.close()
return best_epoch
def main():
parser = argparse.ArgumentParser(description='Training the networks.')
parser.add_argument(
'-lr', '--learning-rate', action='store', dest='learning_rate', required=False, type=float,
default=0.1, help='Learning rate.'
)
parser.add_argument(
'-bs', '--batch-size', action='store', dest='batch_size', required=False, type=int,
default=256, help='Batch size.'
)
parser.add_argument(
'-e', '--epochs', action='store', dest='num_epochs', required=False, type=int,
default=30, help='Number of training epochs.'
)
parser.add_argument(
'-fd', '--feat-dim', action='store', dest='feat_dim', required=False, type=int,
default=64, help='Features dimensionality.'
)
parser.add_argument(
'-fl', '--feat-layer', action='store', dest='feat_layer', required=False, type=str,
default='drop9', help='Features layer.'
)
parser.add_argument(
'-t', '--top', action='store', dest='top', required=False, type=int,
default=10, help='Top <x> for validation.'
)
parser.add_argument(
'-s', '--step-size', action='store', dest='step_size', required=False, type=float,
default=0.33, help='Step size for learning with step-down policy.'
)
parser.add_argument(
'-a', '--activation', action='store', dest='activation', required=False, type=str,
default='sigmoid', help='Activation function (final net layer).'
)
parser.add_argument(
'-sd', '--samples-dir', action='store', dest='samples_dir', required=False, type=str,
default='~/datasets/samples', help='Path where samples (samples) are placed.'
)
parser.add_argument(
'-d', '--dataset', action='store', dest='dataset', required=False, type=str,
default='cdip', help='Dataset id.'
)
parser.add_argument(
'-ma', '--margin', action='store', dest='margin', required=False, type=float,
default=1.0, help='Margin for contrastive divergence loss'
)
parser.add_argument(
'-m', '--model-id', action='store', dest='model_id', required=False, type=str,
default=None, help='Model identifier (tag).'
)
args = parser.parse_args()
# training stage
t0 = time.time()
train(args)
train_time = time.time() - t0
# validation
t0 = time.time()
best_epoch = validate(args)
val_time = time.time() - t0
# dump training info
info = {
'train_time': train_time,
'val_time': val_time,
'sample_height': int(args.samples_dir.split('_')[-1].split('x')[0]),
'best_model_left': 'traindata/{}/model/left/{}.npy'.format(args.model_id, best_epoch),
'best_model_right': 'traindata/{}/model/right/{}.npy'.format(args.model_id, best_epoch),
'params': args.__dict__
}
json.dump(info, open('traindata/{}/info.json'.format(args.model_id), 'w'))
return train_time + val_time
if __name__ == '__main__':
t = main()
print('Elapsed time={:.2f} minutes ({} seconds)'.format(t / 60.0, t))