Can symmetrically paddding be done in convolution layers in Keras?

Question

I read that the padding is same or avlid in convolution layers in Keras, and I think zeros are padded.

Is there any way to do a symmetrically padding in Keras?

It seems that this can be done with TensorFlow's tf.pad. tf.pad(t, paddings, "SYMMETRIC") is just what I want do. Can Keras does that with TensorFlow as the backend?

Answer 1

I've written an example layer in keras which calls the tensorflow padding backend.

import keras.backend as K
from keras.layers import Layer

class SymmetricPadding2D(Layer):

    def __init__(self, output_dim, padding=[1,1], 
                 data_format="channels_last", **kwargs):
        self.output_dim = output_dim
        self.data_format = data_format
        self.padding = padding
        super(SymmetricPadding2D, self).__init__(**kwargs)

    def build(self, input_shape):
        super(SymmetricPadding2D, self).build(input_shape)

    def call(self, inputs):
        if self.data_format is "channels_last":
            #(batch, depth, rows, cols, channels)
            pad = [[0,0]] + [[i,i] for i in self.padding] + [[0,0]]
        elif self.data_format is "channels_first":
            #(batch, channels, depth, rows, cols)
            pad = [[0, 0], [0, 0]] + [[i,i] for i in self.padding]

        if K.backend() == "tensorflow":
            import tensorflow as tf
            paddings = tf.constant(pad)
            out = tf.pad(inputs, paddings, "REFLECT")
        else:
            raise Exception("Backend " + K.backend() + "not implemented")
        return out 

    def compute_output_shape(self, input_shape):
        return (input_shape[0], self.output_dim)

if __name__ == "__main__":

    from keras.models import Sequential
    import numpy as np

    #Set Image
    image = [[1,2,3,4],[5,6,7,8],[9,10,11,12],[13,14,15,16]]

    # Pad to "channels_last format 
    # which is [batch, width, height, channels]=[1,4,4,1]
    image = np.expand_dims(np.expand_dims(np.array(image),2),0)


    #Build Keras model
    model = Sequential()
    model.add(SymmetricPadding2D(1, input_shape=(4,4,1)))
    model.build()

    # To simply apply existing filter, we use predict with no training
    out = model.predict(image)
    print(out[0,:,:,0])