CODEWITHSUNDEEP
pythonkerasconv-neural-networkautoencoder
ablmmcu
ablmmcu

Reputation: 159

Changing decoder output shape in convolutional autoencoder for training

I am trying to train my dataset for the convolutional autoencoder model. I have patched 500,400x400x1 dataset. And my model is shown below.

input_img = Input(shape = (400, 400, 1))

def encoder(input_img):
#encoder
conv1 = Conv2D(4, (3, 3), activation='relu', padding='same')(input_img) 
conv1 = BatchNormalization()(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Conv2D(8, (3, 3), activation='relu', padding='same')(pool1)
conv2 = BatchNormalization()(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2) 
conv3 = Conv2D(16, (3, 3), activation='relu', padding='same')(pool2) 
conv3 = BatchNormalization()(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Conv2D(32, (3, 3), activation='relu', padding='same')(pool3)
conv4 = BatchNormalization()(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Conv2D(64, (3, 3), activation='relu', padding='same')(pool4)
conv5 = BatchNormalization()(conv5)
pool5 = MaxPooling2D(pool_size=(2, 2))(conv5)
conv6 = Conv2D(128, (3, 3), activation='relu', padding='same')(pool5)
conv6 = BatchNormalization()(conv6)
pool6 = MaxPooling2D(pool_size=(2, 2))(conv6)
conv7 = Conv2D(256, (3, 3), activation='relu', padding='same')(pool6) 
conv7 = BatchNormalization()(conv7)
pool7 = MaxPooling2D(pool_size=(2, 2))(conv7)
conv8 = Conv2D(512, (3, 3), activation='relu', padding='same')(pool7)
conv8 = BatchNormalization()(conv8)
return conv8


def decoder(conv8):    
#decoder
conv9 = Conv2D(512, (3, 3), activation='relu', padding='same')(conv8) 
conv9 = BatchNormalization()(conv9)
up1 = UpSampling2D((2,2))(conv9) 
conv12 = Conv2D(256, (3, 3), activation='relu', padding='same')(up1)
conv12 = BatchNormalization()(conv12)
up2 = UpSampling2D((2,2))(conv12) 
conv13 = Conv2D(128, (3, 3), activation='relu', padding='same')(up2) 
conv13 = BatchNormalization()(conv13)
up3 = UpSampling2D((2,2))(conv13) 
conv14 = Conv2D(64, (3, 3), activation='relu', padding='same')(up3)
conv14 = BatchNormalization()(conv14)
up4 = UpSampling2D((2,2))(conv14) 
conv15 = Conv2D(32, (3, 3), activation='relu', padding='same')(up4)
conv15 = BatchNormalization()(conv15)
up5 = UpSampling2D((2,2))(conv15) 
conv16 = Conv2D(16, (3, 3), activation='relu', padding='same')(up5)
conv16 = BatchNormalization()(conv16)
up6 = UpSampling2D((2,2))(conv16) 
conv17 = Conv2D(8, (3, 3), activation='relu', padding='same')(up6) 
conv17 = BatchNormalization()(conv17)
up7 = UpSampling2D((2,2))(conv17) 
conv18 = Conv2D(4, (3, 3), activation='relu', padding='same')(up7) 
conv18 = BatchNormalization()(conv18)
decoded = Conv2D(1, (3,3), activation='relu', padding='same')(conv18)
return decoded

after the encoder process, the shape goes to (3,3,512). But when the decoder process is done, the shape goes up to (384,384,1). And because my input shape is (400,400,1), I can not train the model. Is there any upsampling way to change my decoder (384,384,1) layer to (400,400,1)?

Upvotes: 1

Views: 509

Answers (1)

Stack Player
Stack Player

Reputation: 1490

You can increase the padding in intermediate Con2D layers to make up for the 16 pixels. See this documentation for how to control padding on 2 dimensions.

You can also use explicit padding layers.

Upvotes: 1

Related Questions