-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
181 lines (144 loc) · 7.03 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
'''Train the model'''
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import datetime
import functools
import os
import json
import gc
import tensorflow as tf
import numpy as np
import scipy.misc
import densecap.model as model
import densecap.util as util
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string('train_dir', os.getcwd(), 'Directory with training data (images)')
tf.app.flags.DEFINE_string('region_desc', '', 'Region descriptions file (Visual Genome format)')
tf.app.flags.DEFINE_string('log_dir', os.path.join(os.getcwd(), 'logs'), 'Directory with logs for tensorboard')
tf.app.flags.DEFINE_string('ckpt_dir', os.path.join(os.getcwd(), 'ckpt'), 'Directory for model checkpoints')
tf.app.flags.DEFINE_integer('batch_size', 64, 'Batch size')
tf.app.flags.DEFINE_integer('limit', 0, 'Limit training process to first `limit` files per epoch')
tf.app.flags.DEFINE_integer('epoch', 10, 'Epoch count')
tf.app.flags.DEFINE_integer('log_every', 100, 'Print log messages every `log_every` steps')
tf.app.flags.DEFINE_integer('save_every', 100, 'Save model checkpoint every `save_every` steps')
tf.app.flags.DEFINE_integer('eval_every', 100, 'Eval model every `eval_every` steps')
tf.app.flags.DEFINE_float('learning_rate', 0.01, 'Initial learning rate')
def train_data(filename, limit):
with open(FLAGS.region_desc) as ifile:
data = json.load(ifile)
limit = limit or len(data)
for idx in range(limit):
record = data[idx]
filename = os.path.join(
FLAGS.train_dir, str(record['id']) + '.jpg')
# XXX: resize to have 600/720 longer side
image = scipy.misc.imread(filename, mode='RGB')
height, width, _ = image.shape
fraction = 720.0 / max(height, width)
image = scipy.misc.imresize(image, fraction)
gt_boxes = np.array([[r['y'], r['x'], r['height'], r['width']]
for r in record['regions']])
yield (image, gt_boxes)
def load_vgg16_weights(sess):
with np.load('data/vgg16_weights.npz') as ifile:
for v in tf.global_variables():
name = v.name.replace('weights', 'W').replace('biases', 'b').replace('/', '_')[:-2]
if name in ifile:
sess.run(tf.assign(v, ifile[name]))
def main(_):
'''entry point'''
image_input = tf.placeholder(tf.float32, shape=[1, None, None, 3])
vgg16 = model.VGG16(image_input)
rpn = model.RegionProposalNetwork(vgg16.layers['conv5_3'])
current_run_log_dir = os.path.join(
FLAGS.log_dir,
datetime.datetime.now().isoformat()
)
writer = tf.train.SummaryWriter(current_run_log_dir, graph=tf.get_default_graph())
saver = tf.train.Saver()
saved_model = tf.train.latest_checkpoint(FLAGS.ckpt_dir)
if not os.path.exists(FLAGS.ckpt_dir):
os.makedirs(FLAGS.ckpt_dir)
@functools.lru_cache()
def get_data():
return list(train_data(FLAGS.region_desc, FLAGS.limit))
with tf.Session() as sess:
if saved_model:
saver.restore(sess, saved_model)
else:
print('Prevous model not found, starting from scratch.')
sess.run(tf.global_variables_initializer())
load_vgg16_weights(sess)
for epoch in range(FLAGS.epoch):
for image, gt_boxes in get_data():
height, width, _ = image.shape
p_bbox, p_score, p_label, n_bbox, n_score, n_label = sess.run(
[rpn.positive_bbox, rpn.positive_scores, rpn.positive_labels,
rpn.negative_bbox, rpn.negative_scores, rpn.negative_labels], {
vgg16.input: [image],
rpn.image_height: height,
rpn.image_width: width,
rpn.ground_truth_pre: gt_boxes,
rpn.ground_truth_num: len(gt_boxes)
})
(p_bbox, p_score, p_label), (n_bbox, n_score, n_label) = model.generate_batches(
(p_bbox, p_score, np.expand_dims(p_label, axis=1)),
(n_bbox, n_score, np.expand_dims(n_label, axis=1)),
rpn.batch_size
)
merged = tf.summary.merge_all()
loss, step, summary, _ = sess.run([rpn.loss, rpn.global_step, merged, rpn.train_op], {
vgg16.input: [image],
rpn.image_height: height,
rpn.image_width: width,
rpn.ground_truth_pre: gt_boxes,
rpn.ground_truth_num: len(gt_boxes),
rpn.pos_boxes: p_bbox,
rpn.pos_scores: p_score,
rpn.true_pos_scores: np.squeeze(p_label),
rpn.neg_boxes: n_bbox,
rpn.neg_scores: n_score,
rpn.true_neg_scores: np.squeeze(n_label),
})
writer.add_summary(summary, global_step=step)
gc.collect()
if not step % FLAGS.log_every:
print('\rEpoch {:<2} step {:<6} loss: {:<8.2f}'\
.format(epoch+1, step, loss), end='')
if not step % FLAGS.save_every:
saver.save(
sess,
os.path.join(FLAGS.ckpt_dir, 'densecap'),
global_step=rpn.global_step)
if not step % FLAGS.eval_every:
k = 300
boxes, scores = sess.run(
[rpn.proposals, tf.nn.softmax(rpn.scores)], {
rpn.image_height: height,
rpn.image_width: width,
vgg16.input: [image]
})
np_proposals = np.squeeze(boxes[np.argsort(scores[:, 1])][-k:])
proposals = tf.placeholder(tf.float32, [None, 4])
ground_truth = tf.placeholder(tf.float32, [None, 4])
proposals_num = tf.placeholder(tf.int32)
ground_truth_num = tf.placeholder(tf.int32)
recall, precision = sess.run(
[model.recall(proposals, proposals_num, ground_truth, ground_truth_num, 0.5),
model.precision(proposals, proposals_num, ground_truth, ground_truth_num, 0.5)], {
proposals: np_proposals,
proposals_num: len(np_proposals),
ground_truth: gt_boxes,
ground_truth_num: len(gt_boxes)
})
summary = tf.Summary(value=[
tf.Summary.Value(tag='recall', simple_value=float(recall)),
tf.Summary.Value(tag='precision', simple_value=float(precision)),
])
writer.add_summary(summary, global_step=step)
print()
writer.close()
if __name__ == '__main__':
tf.app.run()