-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodelDef.py
225 lines (161 loc) · 11.5 KB
/
modelDef.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
import keras
from keras.layers import Conv2D, BatchNormalization, MaxPool2D, Flatten, AveragePooling2D, concatenate, Add, Activation, Input, UpSampling2D, Conv2DTranspose
from keras import Model
from keras.optimizers import Adam
from resnetDef import ResNet50, ResNet18
'''
Create a resnet encoder either type 50 layers or type 18 layers...
output 1000 feature values/classes
pad zeros of skill connection from input when channels increase...
main model does a resnet encoding then a depth decoding
the res-net can be tested on and trained on imagenet to verify correctness and working
ref:
https://medium.com/@14prakash/understanding-and-implementing-architectures-of-resnet-and-resnext-for-state-of-the-art-image-cf51669e1624
https://keras.io/applications/#resnet
https://towardsdatascience.com/understanding-and-visualizing-resnets-442284831be8
'''
def generateResNetEncoderLayers(inputLayer, resnetType=18):
'''
takes an input layer of type Input from keras
returns the output layer of a resent of the specifified type
input layer must be of form: (batches, even#, even#, 3*numImages)
'''
assert resnetType in [50,18]
if resnetType == 50:
x = ResNetLayerInitialStage(inputLayer)
x = ResNetLayerLaterState_50(x, 64, [64,64,256],1,1)
x = ResNetLayerLaterState_50(x, 256, [64,64,256],1,2)
x = ResNetLayerLaterState_50(x, 256, [64,64,256],1,3)
x1 = ResNetLayerLaterState_50(x, 256, [128,128,512],2,4)
x1 = ResNetLayerLaterState_50(x1, 512, [128,128,512],1,5)
x1 = ResNetLayerLaterState_50(x1, 512, [128,128,512],1,6)
x1 = ResNetLayerLaterState_50(x1, 512, [128,128,512],1,7)
x2 = ResNetLayerLaterState_50(x1, 512, [256,256,1024],2,8)
x2 = ResNetLayerLaterState_50(x2, 1024, [256,256,1024],1,9)
x2 = ResNetLayerLaterState_50(x2, 1024, [256,256,1024],1,10)
x2 = ResNetLayerLaterState_50(x2, 1024, [256,256,1024],1,11)
x2 = ResNetLayerLaterState_50(x2, 1024, [256,256,1024],1,12)
x2 = ResNetLayerLaterState_50(x2, 1024, [256,256,1024],1,13)
x3 = ResNetLayerLaterState_50(x2, 1024, [512,512,2048],2,14)
x3 = ResNetLayerLaterState_50(x3, 2048, [512,512,2048],1,15)
x3 = ResNetLayerLaterState_50(x3, 2048, [512,512,2048],1,16)
x3 = ResNetOuputStage(x3)
return x3, x2, x1, x
else: # is resnet 18
x = ResNetLayerInitialStage(inputLayer)
x = ResNetLayerLaterState_18(x, 64, [64,64],1,1)
x = ResNetLayerLaterState_18(x, 64, [64,64],1,2)
x1 = ResNetLayerLaterState_18(x, 64, [128,128],2,3)
x1 = ResNetLayerLaterState_18(x1, 128, [128,128],1,4)
x2 = ResNetLayerLaterState_18(x1, 128, [256,256],2,6)
x2 = ResNetLayerLaterState_18(x2, 256, [256,256],1,7)
x3 = ResNetLayerLaterState_18(x2, 256, [512,512],2,8)
x3 = ResNetLayerLaterState_18(x3, 512, [512,512],1,9)
x3 = ResNetOuputStage(x3)
return x3, x2, x1, x
def ResNetLayerInitialStage(inputLayer):
x = Conv2D(filters=64,kernel_size=7,strides=2,data_format='channels_last',activation='relu',padding='same', name="InitialConv")(inputLayer)
x = BatchNormalization(axis=3)(x)
x = MaxPool2D(pool_size=(3,3),strides=2, data_format='channels_last',padding='same', name="InitalPool")(x)
return x
def ResNetLayerLaterState_50(inputLayer, inputChannels, channels, poolingStride, resNetBlockID):
'''
3 convolutional blocks
1x1, channels[0], relu
3x3, channels[1], relu
1x1, channels[2], linear
add input and output
relu
'''
assert len(channels) == 3
x = Conv2D(channels[0], kernel_size=poolingStride, strides=poolingStride,data_format='channels_last',activation='relu',padding='same', name="Conv1_" + str(inputChannels) + "__" + str(resNetBlockID))(inputLayer)
x = BatchNormalization(axis=3)(x)
x = Conv2D(channels[1], kernel_size=3,strides=1,data_format='channels_last',activation='relu',padding='same', name="Conv2_" + str(inputChannels) + "__" + str(resNetBlockID))(x)
x = BatchNormalization(axis=3)(x)
x = Conv2D(channels[2], kernel_size=1,strides=1,data_format='channels_last',activation='linear',padding='same', name="Conv3_" + str(inputChannels) + "__" + str(resNetBlockID))(x)
x = BatchNormalization(axis=3)(x)
if inputChannels != channels[2]:
# this could be zero padding but instread were doing 1x1 stride 1 convolution to make the shapes the same, both are technically from paper acceptable
inputLayer = Conv2D(channels[2], kernel_size=1, strides=1,data_format='channels_last',activation='linear',padding='same', name="ConvSkip_" + str(inputChannels) + "__" + str(resNetBlockID))(inputLayer)
if poolingStride != 1:
inputLayer = MaxPool2D(pool_size=2)(inputLayer)
x = Add()([x,inputLayer])
x = Activation('relu')(x)
return x
def ResNetLayerLaterState_18(inputLayer, inputChannels, channels, poolingStride, resNetBlockID):
'''
two convolutional blocks
3x3, channels[0], relu
3x3, channels[0], linear
add input and output
relu
'''
assert len(channels) == 2
x = Conv2D(channels[0], kernel_size=3,strides=poolingStride,data_format='channels_last',activation='relu',padding='same', name="Conv1_" + str(inputChannels) + "__" + str(resNetBlockID))(inputLayer)
x = BatchNormalization(axis=3)(x)
x = Conv2D(channels[1], kernel_size=3,strides=1,data_format='channels_last',activation='linear',padding='same', name="Conv2_" + str(inputChannels) + "__" + str(resNetBlockID))(x)
x = BatchNormalization(axis=3)(x)
if inputChannels != channels[1]:
# this should be zero padding but just simple one conv for now untill fixed
inputLayer = Conv2D(channels[1], kernel_size=1, strides=1,data_format='channels_last',activation='linear',padding='same', name="ConvSkip_" + str(inputChannels) + "__" + str(resNetBlockID))(inputLayer)
if poolingStride != 1:
inputLayer = MaxPool2D(pool_size=2,padding='same', name="Pool_" + str(inputChannels) + "__" + str(resNetBlockID))(inputLayer)
x = Add()([x,inputLayer])
x = Activation('relu')(x)
return x
def ResNetOuputStage(inputLayer, pools=1000):
output = AveragePooling2D(pool_size=2,strides=1,padding='same')(inputLayer)
return output
def buildDecoder(inputLayer, scale_1, scale_2, scale_3, outputChannels=1):
pass
x = Conv2D(1024, kernel_size=3, strides=1, data_format='channels_last',padding='same', name="DecoderConv_Block_1_1", activation='relu')(inputLayer)
x = Conv2D(1024, kernel_size=3, strides=1, data_format='channels_last',padding='same', name="DecoderConv_Block_1_2", activation='relu')(x)
x = Conv2DTranspose(512, (2, 2), strides=(2, 2), data_format='channels_last', name="ConvTranspose1")(x)
x = concatenate([x,scale_1],axis=3)
x = Conv2D(512, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_2_1", activation='relu')(x)
x = Conv2D(512, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_2_2", activation='relu')(x)
x = Conv2DTranspose(256, (2, 2), strides=(2, 2), data_format='channels_last', name="ConvTranspose2")(x)
x = concatenate([x,scale_2],axis=3)
x = Conv2D(256, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_3_1", activation='relu')(x)
x = Conv2D(256, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_3_2", activation='relu')(x)
scale_3_out = Conv2D(32, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="EndingConvBlock_Scale3")(x)
scale_3_out = Conv2D(outputChannels, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="OutputConvBlock_Scale3")(scale_3_out)
scale_3_out = UpSampling2D(data_format='channels_last', name="upSampleScale3Out", size=(8,8), interpolation='bilinear')(scale_3_out)
x = Conv2DTranspose(128, (2, 2), strides=(2, 2), data_format='channels_last', name="ConvTranspose3")(x)
x = concatenate([x,scale_3],axis=3)
x = Conv2D(128, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_4_1", activation='relu')(x)
x = Conv2D(128, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_4_2", activation='relu')(x)
scale_2_out = Conv2D(32, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="EndingConvBlock_Scale2")(x)
scale_2_out = Conv2D(outputChannels, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="OutputConvBlock_Scale2")(scale_2_out)
scale_2_out = UpSampling2D(data_format='channels_last', name="upSampleScale2Out", size=(4,4), interpolation='bilinear')(scale_2_out)
x = Conv2DTranspose(64, (2, 2), strides=(2, 2), data_format='channels_last', name="ConvTranspose4")(x)
x = Conv2D(64, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_5_1", activation='relu')(x)
x = Conv2D(64, kernel_size=3, strides=1, data_format='channels_last', padding='same', name="DecoderConv_Block_5_2", activation='relu')(x)
scale_1_out = Conv2D(64, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="EndingConvBlock_Scale1")(x)
scale_1_out = Conv2D(outputChannels, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="OutputConvBlock_Scale1")(scale_1_out)
scale_1_out = UpSampling2D(data_format='channels_last', name="upSampleScale1Out", size=(2,2), interpolation='bilinear')(scale_1_out)
x = Conv2DTranspose(64, (2, 2), strides=(2, 2), data_format='channels_last', name="ConvTranspose5")(x)
x = Conv2D(64, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="EndingConvBlock1", activation='relu')(x)
x = Conv2D(64, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="EndingConvBlock2", activation='relu')(x)
x = Conv2D(outputChannels, kernel_size=3, strides=1, data_format='channels_last' ,padding='same', name="OutputConvBlock", activation='sigmoid')(x)
return concatenate([x, scale_1_out, scale_2_out, scale_3_out], axis=3), scale_1_out, scale_2_out, scale_3_out
def create_monoDepth_Model(input_shape=(640,192,3), encoder_type=50):
if encoder_type == 50:
inputLayer, outputLayer, scaleLayers = ResNet50(input_shape=(640,192,3),include_top=False, create_encoder=True)
networkOuput, scale_1_out, scale_2_out, scale_3_out = buildDecoder(outputLayer, scaleLayers[2], scaleLayers[1], scaleLayers[0], 1)
model = Model(inputs=[inputLayer], outputs=[networkOuput])#, scale_1_out, scale_2_out, scale_3_out])
model.load_weights('resnet50_imagenet_1000_no_top.h5', by_name=True)
model.summary()
if encoder_type == 18:
inputLayer, outputLayer, scaleLayers = ResNet18(input_shape=(640,192,3),include_top=False, create_encoder=True)
networkOuput, scale_1_out, scale_2_out, scale_3_out = buildDecoder(outputLayer, scaleLayers[2], scaleLayers[1], scaleLayers[0], 1)
model = Model(inputs=[inputLayer], outputs=[networkOuput])#, scale_1_out, scale_2_out, scale_3_out])
model.load_weights('resnet18_imagenet_1000_no_top.h5', by_name=True)
model.summary()
return model
if __name__ == "__main__":
print("testing creating models")
model = create_monoDepth_Model(input_shape=(640,192,3), encoder_type=18)
print("Done createting ResNet18 backbone model")
model = create_monoDepth_Model(input_shape=(1024,320,3), encoder_type=50)
print("Done createting ResNet50 backbone model")