-
Notifications
You must be signed in to change notification settings - Fork 109
/
resnet.py
executable file
·219 lines (194 loc) · 10.9 KB
/
resnet.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
import logging
import os
import tensorflow as tf
import resnet_v1
import resnet_utils
from resnet_v1 import bottleneck, bottleneck_skip
from utils import print_variables
import config
from config import args, MEAN_COLOR
from paths import INIT_WEIGHTS_DIR
log = logging.getLogger()
slim = tf.contrib.slim
CKPT_50 = os.path.join(INIT_WEIGHTS_DIR, 'resnet50_full.ckpt')
DEFAULT_SCOPE_50 = 'resnet_v1_50'
DEFAULT_SSD_SCOPE = 'ssd'
class ResNet(object):
def __init__(self, config, training, weight_decay=0.0005, depth=50,
reuse=False):
self.config = config
self.weight_decay = weight_decay
self.layers = []
self.reuse = reuse
self.training = training
self.layers = self.config['layers']
if depth == 50:
self.num_block3 = 5
self.scope = DEFAULT_SCOPE_50
self.ckpt = CKPT_50
else:
raise ValueError
def create_trunk(self, images):
red, green, blue = tf.split(images*255, 3, axis=3)
images = tf.concat([blue, green, red], 3) - MEAN_COLOR
with slim.arg_scope(resnet_v1.resnet_arg_scope(is_training=self.training,
weight_decay=self.weight_decay,
batch_norm_decay=args.bn_decay)):
blocks = [
resnet_utils.Block(
'block1', bottleneck, [(256, 64, 1)] * 3),
resnet_utils.Block(
'block2', bottleneck, [(512, 128, 2)] + [(512, 128, 1)] * 3),
resnet_utils.Block(
'block3', bottleneck, [(1024, 256, 2)] + [(1024, 256, 1)] * self.num_block3),
resnet_utils.Block(
'block4', bottleneck, [(2048, 512, 2)] + [(2048, 512, 1)] * 2)
]
net, endpoints = resnet_v1.resnet_v1(images, blocks,
global_pool=False,
reuse=self.reuse,
scope=self.scope)
self.outputs = endpoints
self.add_extra_layers(net)
def vgg_arg_scope(self):
with slim.arg_scope([slim.conv2d],
activation_fn=tf.nn.relu,
weights_regularizer=slim.l2_regularizer(self.weight_decay),
biases_initializer=tf.zeros_initializer(),
padding='SAME') as arg_sc:
return arg_sc
def add_extra_layers(self, net):
with slim.arg_scope(resnet_v1.resnet_arg_scope(is_training=self.training,
weight_decay=self.weight_decay,
batch_norm_decay=args.bn_decay)):
block_depth = 2
num_fm = 2048
blocks = [
resnet_utils.Block(
'block5', bottleneck, [(num_fm, num_fm//4, 2)] + [(num_fm, num_fm//4, 1)] * (block_depth-1)),
resnet_utils.Block(
'block6', bottleneck, [(num_fm, num_fm//4, 2)] + [(num_fm, num_fm//4, 1)] * (block_depth-1)),
resnet_utils.Block(
'block7', bottleneck, [(num_fm, num_fm//4, 2)] + [(num_fm, num_fm//4, 1)] * (block_depth-1)),
]
if args.image_size == 512:
blocks += [
resnet_utils.Block(
'block8', bottleneck, [(num_fm, num_fm//4, 2)] + [(num_fm, num_fm//4, 1)] * (block_depth-1)),
]
net, endpoints = resnet_v1.resnet_v1(net, blocks,
global_pool=False,
include_root_block=False,
reuse=self.reuse,
scope=DEFAULT_SSD_SCOPE)
self.outputs.update(endpoints)
with tf.variable_scope(DEFAULT_SSD_SCOPE+"_back", reuse=self.reuse):
end_points_collection = "reverse_ssd_end_points"
with slim.arg_scope([slim.conv2d, bottleneck_skip],
outputs_collections=end_points_collection):
top_fm = args.top_fm
int_fm = top_fm//4
if args.image_size == 512:
# as long as the number of pooling layers is bigger due to
# the higher resolution, an extra layer is appended
net = bottleneck_skip(net, self.outputs[DEFAULT_SSD_SCOPE+'/block7'],
top_fm, int_fm, scope='block_rev7')
net = bottleneck_skip(net, self.outputs[DEFAULT_SSD_SCOPE+'/block6'],
top_fm, int_fm, scope='block_rev6')
net = bottleneck_skip(net, self.outputs[DEFAULT_SSD_SCOPE+'/block5'],
top_fm, int_fm, scope='block_rev5')
net = bottleneck_skip(net, self.outputs[self.scope+'/block4'],
top_fm, int_fm, scope='block_rev4')
net = bottleneck_skip(net, self.outputs[self.scope+'/block3'],
top_fm, int_fm, scope='block_rev3')
net = bottleneck_skip(net, self.outputs[self.scope+'/block2'],
top_fm, int_fm, scope='block_rev2')
if args.x4:
# To provide stride 4 we add one more layer with upsampling
net = bottleneck_skip(net, self.outputs[self.scope+'/block1'],
top_fm, int_fm, scope='block_rev1')
endpoints = slim.utils.convert_collection_to_dict(end_points_collection)
self.outputs.update(endpoints)
# Creating an output of spatial resolution 1x1 with conventional name 'pool6'
if args.image_size == 512:
self.outputs[DEFAULT_SSD_SCOPE+'/pool6'] =\
tf.reduce_mean(self.outputs['ssd_back/block_rev7/shortcut'],
[1, 2], name='pool6', keep_dims=True)
else:
self.outputs[DEFAULT_SSD_SCOPE+'/pool6'] =\
tf.reduce_mean(self.outputs['ssd_back/block_rev6/shortcut'],
[1, 2], name='pool6', keep_dims=True)
def create_multibox_head(self, num_classes):
"""Creates outputs for classification and localization of all candidate bboxes"""
locations = []
confidences = []
with tf.variable_scope(DEFAULT_SSD_SCOPE, reuse=self.reuse) as sc:
end_points_collection = sc.name + '_end_points'
with slim.arg_scope(self.vgg_arg_scope()):
with slim.arg_scope([slim.conv2d], outputs_collections=end_points_collection,
weights_initializer=slim.variance_scaling_initializer(factor=0.1),
activation_fn=None):
for i, layer_name in enumerate(self.layers):
if i > 0 and args.head == 'shared':
sc.reuse_variables()
if args.head == 'shared':
scope_suffix = ''
elif args.head == 'nonshared':
scope_suffix = '/'+layer_name
else:
raise ValueError
src_layer = self.outputs[layer_name]
shape = src_layer.get_shape()
wh = shape[1] * shape[2]
batch_size = shape[0]
num_priors = len(self.config['aspect_ratios'][i])*2 + 2
loc = slim.conv2d(src_layer, num_priors * 4,
[args.det_kernel, args.det_kernel],
scope='location'+scope_suffix)
loc_sh = tf.stack([batch_size, wh * num_priors, 4])
locations.append(tf.reshape(loc, loc_sh))
tf.summary.histogram("location/"+layer_name, locations[-1])
conf = slim.conv2d(src_layer, num_priors * num_classes,
[args.det_kernel, args.det_kernel],
scope='confidence'+scope_suffix)
conf_sh = tf.stack([batch_size, wh * num_priors, num_classes])
confidences.append(tf.reshape(conf, conf_sh))
tf.summary.histogram("confidence/"+layer_name, confidences[-1])
ssd_end_points = slim.utils.convert_collection_to_dict(end_points_collection)
self.outputs.update(ssd_end_points)
all_confidences = tf.concat(confidences, 1)
all_locations = tf.concat(locations, 1)
self.outputs['location'] = all_locations
self.outputs['confidence'] = all_confidences
return all_confidences, all_locations
def create_segmentation_head(self, num_classes):
"""segmentation of map with stride 8 or 4, if --x4 flag is active"""
with tf.variable_scope(DEFAULT_SSD_SCOPE) as sc:
with slim.arg_scope([slim.conv2d],
kernel_size=args.seg_filter_size,
weights_regularizer=slim.l2_regularizer(self.weight_decay),
biases_initializer=tf.zeros_initializer()):
seg_materials = []
seg_size = self.config['fm_sizes'][0]
for i in range(len(self.layers)):
target_layer = self.outputs[self.layers[i]]
seg = slim.conv2d(target_layer, args.n_base_channels)
seg = tf.image.resize_nearest_neighbor(seg, [seg_size, seg_size])
seg_materials.append(seg)
seg_materials = tf.concat(seg_materials, -1)
seg_logits = slim.conv2d(seg_materials, num_classes,
kernel_size=3, activation_fn=None)
self.outputs['segmentation'] = seg_logits
return self.outputs['segmentation']
def get_imagenet_init(self, opt):
"""optimizer is useful to extract slots corresponding to Adam or Momentum
and exclude them from checkpoint assigning"""
variables = slim.get_variables_to_restore(include=[self.scope])
slots = set()
for v in tf.trainable_variables():
for s in opt.get_slot_names():
slot = opt.get_slot(v, s)
if slot is not None:
slots.add(slot)
variables = list(set(variables) - slots)
return slim.assign_from_checkpoint(self.ckpt, variables) + (variables, )