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
|
import tensorflow as tf
from tensorflow.contrib.slim.python.slim.nets.resnet_v1 import resnet_v1_block, resnet_v1
import tensorflow.contrib.slim as slim
from tensorflow.contrib.slim.python.slim.nets.resnet_utils import arg_scope, conv2d_same
import numpy as np
class FasterRCNNSlim:
def __init__(self):
self._blocks = [resnet_v1_block('block1', base_depth=64, num_units=3, stride=2),
resnet_v1_block('block2', base_depth=128, num_units=4, stride=2),
resnet_v1_block('block3', base_depth=256, num_units=23, stride=1),
resnet_v1_block('block4', base_depth=512, num_units=3, stride=1)]
self._image = tf.placeholder(tf.float32, shape=[1, None, None, 3])
self._im_info = tf.placeholder(tf.float32, shape=[3])
self._anchor_scales = [4, 8, 16, 32]
self._num_scales = len(self._anchor_scales)
self._anchor_ratios = [1]
self._num_ratios = len(self._anchor_ratios)
self._num_anchors = self._num_scales * self._num_ratios
self._scope = 'resnet_v1_101'
with arg_scope([slim.conv2d, slim.conv2d_in_plane, slim.conv2d_transpose, slim.separable_conv2d,
slim.fully_connected],
weights_regularizer=tf.contrib.layers.l2_regularizer(0.0001),
biases_regularizer=tf.no_regularizer,
biases_initializer=tf.constant_initializer(0.0)):
# in _build_network
initializer = tf.random_normal_initializer(stddev=0.01)
initializer_bbox = tf.random_normal_initializer(stddev=0.001)
# in _image_to_head
with slim.arg_scope(self._resnet_arg_scope()):
# in _build_base
with tf.variable_scope(self._scope, self._scope):
net_conv = conv2d_same(self._image, 64, 7, stride=2, scope='conv1')
net_conv = tf.pad(net_conv, [[0, 0], [1, 1], [1, 1], [0, 0]])
net_conv = slim.max_pool2d(net_conv, [3, 3], stride=2, padding='VALID', scope='pool1')
net_conv, _ = resnet_v1(net_conv, self._blocks[:-1], global_pool=False, include_root_block=False,
scope=self._scope)
with tf.variable_scope(self._scope, self._scope):
# in _anchor_component
with tf.variable_scope('ANCHOR-default'):
height = tf.to_int32(tf.ceil(self._im_info[0] / 16.0))
width = tf.to_int32(tf.ceil(self._im_info[1] / 16.0))
shift_x = tf.range(width) * 16
shift_y = tf.range(height) * 16
shift_x, shift_y = tf.meshgrid(shift_x, shift_y)
sx = tf.reshape(shift_x, [-1])
sy = tf.reshape(shift_y, [-1])
shifts = tf.transpose(tf.stack([sx, sy, sx, sy]))
k = width * height
shifts = tf.transpose(tf.reshape(shifts, [1, k, 4]), perm=[1, 0, 2])
anchors = np.array([[-24, -24, 39, 39], [-56, -56, 71, 71],
[-120, -120, 135, 135], [-248, -248, 263, 263]], dtype=np.int32)
a = anchors.shape[0]
anchor_constant = tf.constant(anchors.reshape([1, a, 4]), dtype=tf.int32)
length = k * a
anchors_tf = tf.reshape(anchor_constant + shifts, shape=[length, 4])
anchors = tf.cast(anchors_tf, dtype=tf.float32)
self._anchors = anchors
self._anchor_length = length
# in _region_proposal
rpn = slim.conv2d(net_conv, 512, [3, 3], trainable=False, weights_initializer=initializer,
scope='rpn_conv/3x3')
rpn_cls_score = slim.conv2d(rpn, self._num_anchors * 2, [1, 1], trainable=False,
weights_initializer=initializer, padding='VALID', activation_fn=None,
scope='rpn_cls_score')
rpn_cls_score_reshape = self._reshape(rpn_cls_score, 2, 'rpn_cls_score_reshape')
rpn_cls_prob_reshape = self._softmax(rpn_cls_score_reshape, 'rpn_cls_prob_reshape')
# rpn_cls_pred = tf.argmax(tf.reshape(rpn_cls_score_reshape, [-1, 2]), axis=1, name='rpn_cls_pred')
rpn_cls_prob = self._reshape(rpn_cls_prob_reshape, self._num_anchors * 2, 'rpn_cls_prob')
rpn_bbox_pred = slim.conv2d(rpn, self._num_anchors * 4, [1, 1], trainable=False,
weights_initializer=initializer, padding='VALID', activation_fn=None,
scope='rpn_bbox_pred')
# in _proposal_layer
with tf.variable_scope('rois'):
post_nms_topn = 300
nms_thresh = 0.7
scores = rpn_cls_prob[:, :, :, self._num_anchors:]
scores = tf.reshape(scores, [-1])
rpn_bbox_pred = tf.reshape(rpn_bbox_pred, [-1, 4])
boxes = tf.cast(self._anchors, rpn_bbox_pred.dtype)
widths = boxes[:, 2] - boxes[:, 0] + 1.0
heights = boxes[:, 3] - boxes[:, 1] + 1.0
ctr_x = boxes[:, 0] + widths * 0.5
ctr_y = boxes[:, 1] + heights * 0.5
dx = rpn_bbox_pred[:, 0]
dy = rpn_bbox_pred[:, 1]
dw = rpn_bbox_pred[:, 2]
dh = rpn_bbox_pred[:, 3]
pred_ctr_x = dx * widths + ctr_x
pred_ctr_y = dy * heights + ctr_y
pred_w = tf.exp(dw) * widths
pred_h = tf.exp(dh) * heights
pred_boxes0 = pred_ctr_x - pred_w * 0.5
pred_boxes1 = pred_ctr_y - pred_h * 0.5
pred_boxes2 = pred_ctr_x + pred_w * 0.5
pred_boxes3 = pred_ctr_y + pred_h * 0.5
b0 = tf.clip_by_value(pred_boxes0, 0, self._im_info[1] - 1)
b1 = tf.clip_by_value(pred_boxes1, 0, self._im_info[0] - 1)
b2 = tf.clip_by_value(pred_boxes2, 0, self._im_info[1] - 1)
b3 = tf.clip_by_value(pred_boxes3, 0, self._im_info[0] - 1)
proposals = tf.stack([b0, b1, b2, b3], axis=1)
indices = tf.image.non_max_suppression(proposals, scores, max_output_size=post_nms_topn,
iou_threshold=nms_thresh)
boxes = tf.to_float(tf.gather(proposals, indices))
# rpn_scores = tf.reshape(tf.gather(scores, indices), [-1, 1])
batch_inds = tf.zeros([tf.shape(indices)[0], 1], dtype=tf.float32)
rois = tf.concat([batch_inds, boxes], 1)
# in _crop_pool_layer
with tf.variable_scope('pool5'):
batch_ids = tf.squeeze(tf.slice(rois, [0, 0], [-1, 1], name='bath_id'), [1])
bottom_shape = tf.shape(net_conv)
height = (tf.to_float(bottom_shape[1]) - 1) * 16.0
width = (tf.to_float(bottom_shape[2]) - 1) * 16.0
x1 = tf.slice(rois, [0, 1], [-1, 1], name='x1') / width
y1 = tf.slice(rois, [0, 2], [-1, 1], name='y1') / height
x2 = tf.slice(rois, [0, 3], [-1, 1], name='x2') / width
y2 = tf.slice(rois, [0, 4], [-1, 1], name='y2') / height
bboxes = tf.stop_gradient(tf.concat([y1, x1, y2, x2], 1))
pool5 = tf.image.crop_and_resize(net_conv, bboxes, tf.to_int32(batch_ids), [7, 7], name='crops')
# in _head_to_tail
with slim.arg_scope(self._resnet_arg_scope()):
fc7, _ = resnet_v1(pool5, self._blocks[-1:], global_pool=False, include_root_block=False,
scope=self._scope)
fc7 = tf.reduce_mean(fc7, axis=[1, 2])
with tf.variable_scope(self._scope, self._scope):
# in _region_classification
cls_score = slim.fully_connected(fc7, 2, weights_initializer=initializer, trainable=False,
activation_fn=None, scope='cls_score')
cls_prob = self._softmax(cls_score, 'cls_prob')
# cls_pred = tf.argmax(cls_score, 'cls_pred')
bbox_pred = slim.fully_connected(fc7, 2*4, weights_initializer=initializer_bbox, trainable=False,
activation_fn=None, scope='bbox_pred')
self._cls_score = cls_score
self._cls_prob = cls_prob
self._bbox_pred = bbox_pred
self._rois = rois
stds = np.tile(np.array([0.1, 0.1, 0.2, 0.2]), 2)
means = np.tile(np.array([0.0, 0.0, 0.0, 0.0]), 2)
self._bbox_pred *= stds
self._bbox_pred += means
@staticmethod
def _resnet_arg_scope():
batch_norm_params = {
'is_training': False,
'decay': 0.997,
'epsilon': 1e-5,
'scale': True,
'trainable': False,
'updates_collections': tf.GraphKeys.UPDATE_OPS
}
with arg_scope([slim.conv2d],
weights_regularizer=slim.l2_regularizer(0.0001),
weights_initializer=slim.variance_scaling_initializer(),
trainable=False,
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params):
with arg_scope([slim.batch_norm], **batch_norm_params) as arg_sc:
return arg_sc
@staticmethod
def _reshape(bottom, num_dim, name):
input_shape = tf.shape(bottom)
with tf.variable_scope(name):
to_caffe = tf.transpose(bottom, [0, 3, 1, 2])
reshaped = tf.reshape(to_caffe, [1, num_dim, -1, input_shape[2]])
to_tf = tf.transpose(reshaped, [0, 2, 3, 1])
return to_tf
@staticmethod
def _softmax(bottom, name):
if name.startswith('rpn_cls_prob_reshape'):
input_shape = tf.shape(bottom)
bottom_reshaped = tf.reshape(bottom, [-1, input_shape[-1]])
reshaped_score = tf.nn.softmax(bottom_reshaped, name=name)
return tf.reshape(reshaped_score, input_shape)
return tf.nn.softmax(bottom, name=name)
def test_image(self, sess, image, im_info):
return sess.run([self._cls_score, self._cls_prob, self._bbox_pred, self._rois], feed_dict={
self._image: image,
self._im_info: im_info
})
|
