Keras autoencoder outputting wrong shape

Question

I'm trying to build a deep convolutional autoencoder in Keras, but it keeps outputting the wrong shape.

Code:

def build_network(input_shape):
    input_input =  Input(shape=input_shape)

    #Encode
    x = Conv2D(16, (3, 3), activation='relu', padding = 'same')(input_input)
    x = MaxPooling2D((2, 2), padding='same')(x)
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = MaxPooling2D((2, 2), padding='same')(x)
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = MaxPooling2D((2, 2), padding='same')(x)

    #Decode
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = UpSampling2D((2, 2))(x)
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = UpSampling2D((2, 2))(x)
    x = Conv2D(16, (3, 3), activation='relu', padding='same')(x) 
    x = UpSampling2D((2, 2))(x)
    decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)
    autoencoder = Model(input_input, decoded)
    return autoencoder


if __name__ == "__main__":
    print(build_network((1, 32, 32)).layers[-1].output)

I expect the output shape to be the same as the input shape, but it is instead (8, 32, 1) for (1, 32, 32)

Krunal V Krunal V · Accepted Answer · 2019-04-25T13:46:36

Try to use print(build_network((32,32,1)).layers[-1].output). Or if you want to use channel first than you need to change model like this,

def build_network(input_shape):
    input_input =  Input(shape=input_shape)

    #Encode
    x = Conv2D(16, (3, 3), activation='relu', padding = 'same')(input_input)
    x = MaxPooling2D((2, 2), padding='same')(x)
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = MaxPooling2D((2, 2), padding='same')(x)
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = MaxPooling2D((2, 2), padding='same')(x)

    #Decode
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = UpSampling2D(size=(2, 2),data_format="channels_first")(x)
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    x = UpSampling2D(size=(2, 2),data_format="channels_first")(x)
    x = Conv2D(16, (3, 3), activation='relu', padding='same')(x) 
    decoded = UpSampling2D(size=(2, 2),data_format="channels_first")(x)
    # decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)
    autoencoder = Model(input_input, decoded)
    return autoencoder

if __name__ == "__main__":
    print(build_network((1, 32, 32)).layers[-1].output)

Because in UpSampling2D, default is "channels_last".

Keras autoencoder outputting wrong shape

1 Answers