def createNewModel(patchSize): seed=5 np.random.seed(seed) input=Input(shape=(1,patchSize[0, 0], patchSize[0, 1])) out1=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(input) out2=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out1) sout3=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out2) out3=Conv2D(filters=64,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out2) out4=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out3) out4=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out4=add([sout3,out4]) out4=pool2(pool_size=(2,2),data_format='channels_first')(out4) sout5=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out5=Conv2D(filters=64,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out6=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5) out6=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out6) out6=add([sout5,out6]) out6=pool2(pool_size=(2,2),data_format='channels_first')(out6) out10=Flatten()(out6) out11=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out10) cnn = Model(inputs=input,outputs=out11) return cnn
def createNewModel(patchSize): seed=5 np.random.seed(seed) input=Input(shape=(1,patchSize[0, 0], patchSize[0, 1])) out=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(input) out1=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out) #out1=pool2(pool_size=(3,3),strides=(2,2),data_format='channels_first')(out) sout1=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out1) out=Block(out1,128,with_shortcut=True) out=Block(out,128,with_shortcut=False) out=add([sout1,out]) out=concatenate(inputs=[out1,out],axis=1) out2=pool2(pool_size=(3,3),strides=(2,2),data_format='channels_first')(out) sout2=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out2) out=Block(out2,256,with_shortcut=True) out=Block(out,256,with_shortcut=False) out=add([sout2,out]) out1=pool2(pool_size=(3,3),strides=(2,2),data_format="channels_first")(out1) out=concatenate(inputs=[out1,out2,out],axis=1) out3=pool2(pool_size=(3,3),strides=(2,2),data_format='channels_first')(out) #out5=GlobalAveragePooling2D(data_format='channels_first')(out3) out5=Flatten()(out3) out6=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out5) cnn = Model(inputs=input,outputs=out6) return cnn
def createNewModel(patchSize): seed=5 np.random.seed(seed) input=Input(shape=(1,patchSize[0, 0], patchSize[0, 1])) out1=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(input) out2=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out1) out2=pool2(pool_size=(2,2),data_format='channels_first')(out2) out3=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out2) out4=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out3) out4=add([out2,out4]) out4=pool2(pool_size=(2,2),data_format='channels_first')(out4) out5_1=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out5_2=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out5_2=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_2) out5_3=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out5_3=Conv2D(filters=128,kernel_size=(5,5),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_3) out5_4=pool2(pool_size=(3,3),strides=(1,1),padding='same',data_format='channels_first')(out4) out5_4=Conv2D(filters=128,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_4) out5=concatenate(inputs=[out5_1,out5_2,out5_3],axis=1) # out6_1=Conv2D(filters=32,kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5) # # out6_2=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5) # out6_2=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out6_2) # # out6_3=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5) # out6_3=Conv2D(filters=128,kernel_size=(5,5),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out6_3) # # out6_4=pool2(pool_size=(3,3),strides=(1,1),padding='same',data_format='channels_first')(out5) # out6_4=Conv2D(filters=128,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out6_4) # # out6=concatenate(inputs=[out6_1,out6_2,out6_3,out6_4],axis=1) out7=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6), activation='relu')(out5) out8=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6), activation='relu')(out7) out8=add([out5, out8]) out8=pool2(pool_size=(2,2),data_format='channels_first')(out8) out9=Flatten()(out8) out10=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out9) cnn = Model(inputs=input,outputs=out10) return cnn
def create180180Model(patchSize, iVersion=1): seed = 5 np.random.seed(seed) input = Input(shape=(1, patchSize[0, 0], patchSize[0, 1])) # DenseResNet 180180 (selected) if (iVersion == 1): out = Conv2D(filters=64, kernel_size=(3, 3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(2, 2), kernel_regularizer=l2(1e-6), activation='relu')(input) out = Conv2D(filters=64, kernel_size=(3, 3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out) out1 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out1, 128, with_shortcut=True) out = Block(out, 128, with_shortcut=False) out = concatenate(inputs=[out1, out], axis=1) out2 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out2, 128, with_shortcut=True) out = Block(out, 128, with_shortcut=False) out1 = pool2(pool_size=(2, 2), data_format="channels_first")(out1) out = concatenate(inputs=[out1, out2, out], axis=1) out3 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out3, 256, with_shortcut=True) out = Block(out, 256, with_shortcut=False) out2 = pool2(pool_size=(2, 2), data_format="channels_first")(out2) out = concatenate(inputs=[out2, out3, out], axis=1) out4 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) #out5=GlobalAveragePooling2D(data_format='channels_first')(out4) out5 = Flatten()(out4) out6 = Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out5) cnn = Model(inputs=input, outputs=out6) # temp/resdensenet2 for 180180 elif iVersion == 2: out = Conv2D(filters=64, kernel_size=(3, 3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(input) out = Conv2D(filters=64, kernel_size=(3, 3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out) out1 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out1, 128, with_shortcut=True) out = Block(out, 128, with_shortcut=False) out = Block(out, 128, with_shortcut=False) out = concatenate(inputs=[out1, out], axis=1) out2 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out2, 128, with_shortcut=True) out = Block(out, 128, with_shortcut=False) out = Block(out, 128, with_shortcut=False) out1 = pool2(pool_size=(2, 2), data_format="channels_first")(out1) out = concatenate(inputs=[out1, out2, out], axis=1) out3 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out3, 256, with_shortcut=True) out = Block(out, 256, with_shortcut=False) out = Block(out, 256, with_shortcut=False) out2 = pool2(pool_size=(2, 2), data_format="channels_first")(out2) out = concatenate(inputs=[out2, out3, out], axis=1) out4 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) # out5=GlobalAveragePooling2D(data_format='channels_first')(out4) out5 = Flatten()(out4) out6 = Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out5) cnn = Model(inputs=input, outputs=out6) elif iVersion == 3: out = Conv2D(filters=64, kernel_size=(3, 3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(input) out = Conv2D(filters=64, kernel_size=(3, 3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out) out1 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out1, 128, with_shortcut=True) out = Block(out, 128, with_shortcut=False) out = Block(out, 128, with_shortcut=False) out = concatenate(inputs=[out1, out], axis=1) out2 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out2, 128, with_shortcut=True) out = Block(out, 128, with_shortcut=False) out = Block(out, 128, with_shortcut=False) out1 = pool2(pool_size=(3, 3), strides=(2, 2), data_format="channels_first")(out1) out = concatenate(inputs=[out1, out2, out], axis=1) out3 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) out = Block(out3, 256, with_shortcut=True) out = Block(out, 256, with_shortcut=False) out2 = pool2(pool_size=(3, 3), strides=(2, 2), data_format="channels_first")(out2) out = concatenate(inputs=[out2, out3, out], axis=1) out4 = pool2(pool_size=(3, 3), strides=(2, 2), data_format='channels_first')(out) # out5=GlobalAveragePooling2D(data_format='channels_first')(out4) out5 = Flatten()(out4) out6 = Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out5) cnn = Model(inputs=input, outputs=out6) return cnn
def createModel(patchSize, iVersion = 1): seed=5 np.random.seed(seed) input=Input(shape=(1,patchSize[0, 0], patchSize[0, 1])) # googlenet7 for 120120 180180 if iVersion == 1: out1=Conv2D(filters=32, kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(input) out2=Conv2D(filters=64, kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out1) out2=pool2(pool_size=(2,2),data_format='channels_first')(out2) out3=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out2) out4=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out3) out4=pool2(pool_size=(2,2),data_format='channels_first')(out4) out5_1=Conv2D(filters=32, kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out4) out5_2=Conv2D(filters=32, # learning rate: 0.1 -> 76% kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out4) out5_2=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5_2) out5_3=Conv2D(filters=32, # learning rate: 0.1 -> 76% kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out4) out5_3=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(5,5), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5_3) out5_4=pool2(pool_size=(3,3),strides=(1,1),padding='same',data_format='channels_first')(out4) out5_4=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5_4) out5=concatenate(inputs=[out5_1,out5_2,out5_3],axis=1) out7=Conv2D(filters=256, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5) #out7=pool2(pool_size=(2,2),data_format='channels_first')(out7) out8=Conv2D(filters=256, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out7) out8=pool2(pool_size=(2,2),data_format='channels_first')(out8) out9=Flatten()(out8) out10=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out9) cnn = Model(inputs=input,outputs=out10) #336C2P for 4040 elif iVersion == 2: out1=Conv2D(filters=32, kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(input) out2=Conv2D(filters=64, kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out1) out3=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out2) out4=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out3) out4=pool2(pool_size=(2,2),data_format='channels_first')(out4) out5=Conv2D(filters=256, kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out4) out6=Conv2D(filters=256, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5) out6=pool2(pool_size=(2,2),data_format='channels_first')(out6) out7=Flatten()(out6) out10=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out7) cnn = Model(inputs=input,outputs=out10) #temp/googlenet764 elif iVersion == 3: out1=Conv2D(filters=64, kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(input) out2=Conv2D(filters=64, kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out1) out2=pool2(pool_size=(2,2),data_format='channels_first')(out2) out3=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out2) out4=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out3) out4=pool2(pool_size=(2,2),data_format='channels_first')(out4) out5_1=Conv2D(filters=32, kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out4) out5_2=Conv2D(filters=32, # learning rate: 0.1 -> 76% kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out4) out5_2=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5_2) out5_3=Conv2D(filters=32, # learning rate: 0.1 -> 76% kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out4) out5_3=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(5,5), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5_3) out5_4=pool2(pool_size=(3,3),strides=(1,1),padding='same',data_format='channels_first')(out4) out5_4=Conv2D(filters=128, # learning rate: 0.1 -> 76% kernel_size=(1,1), kernel_initializer='he_normal', weights=None, padding='same', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5_4) out5=concatenate(inputs=[out5_1,out5_2,out5_3],axis=1) out7=Conv2D(filters=256, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out5) #out7=pool2(pool_size=(2,2),data_format='channels_first')(out7) out8=Conv2D(filters=256, # learning rate: 0.1 -> 76% kernel_size=(3,3), kernel_initializer='he_normal', weights=None, padding='valid', strides=(1, 1), kernel_regularizer=l2(1e-6), activation='relu')(out7) out8=pool2(pool_size=(2,2),data_format='channels_first')(out8) out9=Flatten()(out8) out10=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out9) cnn = Model(inputs=input,outputs=out10) #temp/resgooglenet764 wrong!!! elif iVersion == 4: out1=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(input) out2=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out1) out2=pool2(pool_size=(2,2),data_format='channels_first')(out2) out3=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out2) out4=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out3) out4=add([out2,out4]) out4=pool2(pool_size=(2,2),data_format='channels_first')(out4) out5_1=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out5_2=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out5_2=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_2) out5_3=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out5_3=Conv2D(filters=128,kernel_size=(5,5),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_3) out5_4=pool2(pool_size=(3,3),strides=(1,1),padding='same',data_format='channels_first')(out4) out5_4=Conv2D(filters=128,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_4) out5=concatenate(inputs=[out5_1,out5_2,out5_3],axis=1) # out6_1=Conv2D(filters=32,kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5) # # out6_2=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5) # out6_2=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out6_2) # # out6_3=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5) # out6_3=Conv2D(filters=128,kernel_size=(5,5),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out6_3) # # out6_4=pool2(pool_size=(3,3),strides=(1,1),padding='same',data_format='channels_first')(out5) # out6_4=Conv2D(filters=128,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out6_4) # # out6=concatenate(inputs=[out6_1,out6_2,out6_3,out6_4],axis=1) out7=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6), activation='relu')(out5) out8=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6), activation='relu')(out7) out8=add([out5, out8]) out8=pool2(pool_size=(2,2),data_format='channels_first')(out8) out9=Flatten()(out8) out10=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out9) cnn = Model(inputs=input,outputs=out10) #temp/resgooglenet764_v1 elif iVersion == 5: out1=Conv2D(filters=32,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(input) out2=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='valid',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out1) out2=pool2(pool_size=(2,2),data_format='channels_first')(out2) out3=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out2) out4=Conv2D(filters=64,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out3) out4=add([out2,out4]) out4=pool2(pool_size=(2,2),data_format='channels_first')(out4) out_3=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out4) out_4=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out_3) out5_1=Conv2D(filters=64,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out_4) out5_2=Conv2D(filters=32,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out_4) out5_2=Conv2D(filters=128,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_2) out5_3=Conv2D(filters=64,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out_4) out5_3=Conv2D(filters=128,kernel_size=(5,5),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_3) out5_4=pool2(pool_size=(3,3),strides=(1,1),padding='same',data_format='channels_first')(out4) out5_4=Conv2D(filters=128,kernel_size=(1,1),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6),activation='relu')(out5_4) out5=concatenate(inputs=[out5_1,out5_2,out5_3],axis=1) sout6=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6), activation='relu')(out5) out7=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6), activation='relu')(out5) out8=Conv2D(filters=256,kernel_size=(3,3),kernel_initializer='he_normal',weights=None,padding='same',strides=(1, 1),kernel_regularizer=l2(1e-6), activation='relu')(out7) out8=add([sout6, out8]) out8=pool2(pool_size=(2,2),data_format='channels_first')(out8) out9=Flatten()(out8) out10=Dense(units=11, kernel_initializer='normal', kernel_regularizer='l2', activation='softmax')(out9) cnn = Model(inputs=input,outputs=out10) return cnn