-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlayers.py
214 lines (174 loc) · 7.94 KB
/
layers.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
import keras.engine.topology as KE
import keras.layers as KL
import tensorflow as tf
import numpy as np
from math import e
class proposal(KE.Layer):
def __init__(self, nms_thresh, proposal_count, imagesize, **kwargs):
super(proposal,self).__init__(**kwargs)
self.imagesize = imagesize
self.nms_thresh = nms_thresh
self.proposal_count = proposal_count
def call(self, inputs):
#input[rpn2K,rpn4K,anchorsCWH]
frontScore = inputs[0][:,:,1]
idxs = tf.squeeze(tf.where(tf.greater(frontScore[0],0.7)),1)
deltes = inputs[1][:]
deltes = tf.gather(deltes[0],idxs)
frontScore = tf.gather(frontScore[0],idxs)
anchors = inputs[2][:]
anchors = tf.gather(anchors[0],idxs)
prenms_num = tf.minimum(2000,tf.shape(anchors)[0])
idxs = tf.nn.top_k(frontScore,prenms_num).indices
frontScore = tf.gather(frontScore,idxs)
deltes = tf.gather(deltes,idxs)
anchors = tf.gather(anchors,idxs)
deltaX = deltes[:,0]
deltaY = deltes[:,1]
deltaW = deltes[:,2]
deltaH = deltes[:,3]
Pw = (e**deltaW)*anchors[:,2]
Ph = (e**deltaH)*anchors[:,3]
Px = (deltaX*anchors[:,2])+anchors[:,0]
Py = (deltaX*anchors[:,3])+anchors[:,1]
refined_anchorsCWH = tf.stack([Px, Py, Pw, Ph],1)
center = refined_anchorsCWH[:,0:2]
WH = refined_anchorsCWH[:,2:4]
refined_anchorsSE = tf.concat([center-WH*0.5,center+WH*0.5],1)
#NMS?
idxs = tf.image.non_max_suppression(refined_anchorsSE,frontScore,25, 0.3)
box = tf.cast(tf.gather(refined_anchorsSE, idxs),tf.int32)
boxxs = tf.clip_by_value(box[:,0], 0,self.imagesize[1])
boxxe = tf.clip_by_value(box[:,2], 0,self.imagesize[1])
boxys = tf.clip_by_value(box[:,1], 0,self.imagesize[0])
boxye = tf.clip_by_value(box[:,3], 0,self.imagesize[0])
nms = tf.expand_dims(tf.stack([boxxs,boxys,boxxe,boxye],1),0)
'''
self.outputshape = tf.shape(nms)[0]
'''
return [nms,tf.expand_dims(self.imagesize,0)]
#def compute_output_shape(self, input_shape):#？
#return (inputs, 10)
class target_detector(KE.Layer):
def __init__(self, imagesize, **kwargs):
#[info['bbox']] x 2
super(target_detector,self).__init__(**kwargs)
self.imagesize = imagesize
def IOU(self,bbox,anchors):
xs1,ys1,xe1,ye1 = bbox[0],bbox[1],bbox[2],bbox[3]
xs2,ys2,xe2,ye2 = anchors[:,0],anchors[:,1],anchors[:,2],anchors[:,3]
xs = tf.maximum(xs1,xs2)
xe = tf.minimum(xe1,xe2)
ys = tf.maximum(ys1,ys2)
ye = tf.minimum(ye1,ye2)
xo = tf.maximum(xe-xs,0)
yo = tf.maximum(ye-ys,0)
overarea = xo*yo
#xymul = (xs1-xe1)*(ys1-ye1)+(xs2-xe2)*(ys2-ye2)-overarea
xymul = (xs1-xe1)*(ys1-ye1)#+(xs2-xe2)*(ys2-ye2)-overarea
IOU = overarea / xymul
return IOU
def deltas(self, gt_bboxs, proposal, imgsize, classID):
#gtbox N x 4 CWH
#classID N
#proposal N x 4 SE
Ax = (proposal[:,2]+proposal[:,0])/2
Ay = (proposal[:,3]+proposal[:,1])/2
Aw = tf.abs(proposal[:,2]-proposal[:,0])
Ah = tf.abs(proposal[:,3]-proposal[:,1])
bboxs = tf.gather(gt_bboxs,classID)
deltaX = (bboxs[:,0]-Ax)/Aw
deltaY = (bboxs[:,1]-Ay)/Ah
deltaW = tf.log(bboxs[:,2]/Aw)
deltaH = tf.log(bboxs[:,3]/Ah)
deltas = tf.stack([deltaX, deltaY, deltaW, deltaH],1)
'''
proposal = tf.expand_dims(proposal,1)
proposal = tf.concat([proposal for i in range(len(gt_bboxs))],1)
imgS = tf.concat([imgsize,imgsize],0)
sizeDevidend = tf.stack([imgS for i in range(len(gt_bboxs))],1)
deltes = -(proposal-gt_bboxs)/sizeDevidend
deltes = tf.gather(deltes,classID)
deltes /=0.01
'''
'''
ps = proposal[:,0:2]
pe = proposal[:,2:4]
pc = (ps+pe)/2
pwh = pe-ps
proposalCWH = tf.concat([pc,pwh],1)
shape = tf.shape(proposal)[0]
proposal_ID = tf.cast(tf.expand_dims(tf.range(0,shape),-1),tf.int32)
cla = tf.cast(tf.expand_dims(classID,-1),tf.int32)
imgS = tf.stack([imgsize[1],imgsize[0]],0)
k = tf.cast(tf.ones((tf.shape(proposal)[0],2)),tf.int32)
imgDicidend = imgS * k
WHDividend = tf.cast(tf.gather(gt_bboxs,classID)[:,2:],tf.int32)
dividend = tf.concat([imgDicidend,WHDividend],1)
deltas = proposalCWH/tf.cast(dividend,tf.float32)
deltas /= 0.01
'''
return deltas
def call(self, inputs):
#input[proposal,gt_bboxsCWH]
proposal = tf.cast(inputs[0],tf.float32)
gt_bboxsCWH1 = tf.concat([[[[1.0,1.0,1.0,1.0,0.0]]],inputs[1]],1)
gt_bboxsCWH = gt_bboxsCWH1[:,:,:4]#gt N x 5 [x1,x2,y1,y2,classid]
center = gt_bboxsCWH[:,:,0:2]
WH = gt_bboxsCWH[:,:,2:4]
gt_bboxsSE = tf.cast(tf.concat([center-WH/2,center+WH/2],2),tf.float32)
gt_class = gt_bboxsCWH1[:,:,-1][0]
IOU = tf.transpose(tf.map_fn(lambda x:self.IOU(x,proposal[0]),gt_bboxsSE[0],tf.float32))
IOU = tf.slice(IOU,[0,1],[-1,-1])
bk = tf.fill([tf.shape(IOU)[0],1],0.1)
IOU = tf.concat([bk,IOU],1)
classIdxs = tf.argmax(IOU, axis=1)
classID = tf.gather(gt_class,classIdxs)
deltas = self.deltas(gt_bboxsCWH[0], proposal[0],self.imagesize, classIdxs)
''''''
#self.outputshape = tf.shape(proposal)[0]
return [tf.expand_dims(classID,0), tf.expand_dims(deltas,0)]
class classifier(KE.Layer):
def __init__(self, poolingsize, fp_size, filters, class_num, **kwargs):
super(classifier,self).__init__(**kwargs)
self.poolingsize = (poolingsize*2,poolingsize*2)
self.fp_size = fp_size
self.filters = filters
self.class_num = class_num
def roipooling(self, featuremap, proposal):
x1 = proposal[0][:,0]
y1 = proposal[0][:,1]
x2 = proposal[0][:,2]
y2 = proposal[0][:,3]
proposal = tf.cast(tf.stack([y1/self.fp_size[0],x1/self.fp_size[1],y2/self.fp_size[0],x2/self.fp_size[1]],1),tf.float32)
shape = tf.shape(proposal)[0]
shape = tf.zeros(shape,dtype=tf.int32)
featuremap = tf.image.crop_and_resize(featuremap,proposal,shape,self.poolingsize)
''''''
return featuremap
def call(self, inputs):
#[featuremap, proposal]
featuremap = inputs[0]
proposal = inputs[1]
filter = self.filters
roipooling1 = self.roipooling(featuremap, proposal)
roipooling = KL.pooling.MaxPool2D()(roipooling1)
roipooling = KL.Conv2D(filter,(7,7),padding='valid',strides=2)(roipooling)
roipooling = KL.BatchNormalization(axis=3)(roipooling)
roipooling = KL.Activation('relu')(roipooling)
roipooling = KL.core.Flatten()(roipooling)
roipooling = KL.Conv2D(4096,(1,1),padding='valid',strides=2)(roipooling)
roipooling = KL.BatchNormalization(axis=3)(roipooling)
roipooling = KL.Activation('relu')(roipooling)
roipooling = KL.Conv2D(4096,(1,1),padding='valid',strides=2)(roipooling)
roipooling = KL.BatchNormalization(axis=3)(roipooling)
roipooling = KL.Activation('relu')(roipooling)
Fclass = KL.Conv2D(self.class_num,(1,1))(roipooling)
Fclass = KL.BatchNormalization(axis=3)(Fclass)
Fclass = KL.Activation('softmax')(Fclass)
Fclass = KL.Lambda(lambda x:tf.reshape(x,[1,-1,self.class_num]))(Fclass)
Fdeltes = KL.Conv2D(4*self.class_num,(1,1))(roipooling)
Fdeltes = KL.BatchNormalization(axis=3)(Fdeltes)
Fdeltes = KL.Activation('linear')(Fdeltes)
Fdeltes = KL.Lambda(lambda x:tf.reshape(x,[1,-1,self.class_num,4]))(Fdeltes)
return [Fclass,Fdeltes]