Un-scaled version of CoordConv2D

coordconv2d.py

import keras.backend as K
import tensorflow as tf
from keras.layers import Layer
from keras.layers import Conv2D

"""Not tested, I'll play around with GANs soon with it."""


class CoordConv2D(Layer):
    def __init__(self, channel, kernel, padding='valid', activation=None, **kwargs):
        self.layer = Conv2D(channel, kernel, padding=padding, activation=activation)
        self._name = 'CoordConv2D'
        super(CoordConv2D, self).__init__(**kwargs)

    def call(self, input):
        input_shape = tf.unstack(K.shape(input))
        if K.image_data_format() == 'channel_first':
            bs, channel, w, h = input_shape
        else:
            bs, w, h, channel = input_shape

        # Get indices
        indices = tf.cast(tf.where(K.ones([bs, w, h])), dtype=tf.float32)
        canvas = K.reshape(indices, [bs, w, h, 3])[..., 1:]
        # Normalize the canvas
        canvas = canvas / tf.cast(K.reshape([w, h], [1, 1, 1, 2]), dtype=tf.float32)
        canvas = (canvas * 2) - 1

        # If channel_first, we swap
        if K.image_data_format() == 'channel_first':
            canvas = K.swap_axes(canvas, [0, 3, 1, 2])

        # Concatenate channel-wise
        input = K.concatenate([input, canvas], -1)
        return self.layer(input)

    def compute_output_shape(self, input_shape):
        return self.layer.compute_output_shape(input_shape)


from keras.layers import Input, Conv2D
import numpy as np
from keras import Model

inp = Input([32, 32, 3])
layer = CoordConv2D(63, 3, padding='same')
x = layer(inp)
mod = Model(inp, x)
mod.compile('sgd', 'mse')
res = mod.predict(np.ones([3, 32, 32, 3]))
print(res.shape)