Coordconv as seen in https://arxiv.org/abs/1807.03247

coordconv.py

import tensorflow as tf


def add_coord_channels(inputs, with_r=False):
    batch_size_tensor = tf.shape(inputs)[0]
    x_dim, y_dim = inputs.shape[1].value, inputs.shape[2].value
    x_range, y_range = (tf.linspace(-1., 1., d) for d in (x_dim, y_dim))

    x_channel = tf.tile(tf.expand_dims(y_range, 0), [x_dim, 1])
    y_channel = tf.tile(tf.expand_dims(x_range, 1), [1, y_dim])
    x_channel, y_channel = (tf.expand_dims(tf.expand_dims(x, -1), 0) for x in (x_channel, y_channel))
    x_channel, y_channel = (tf.tile(x, [batch_size_tensor, 1, 1, 1]) for x in (x_channel, y_channel))
    res = tf.concat([inputs, x_channel, y_channel], axis=-1)
    if with_r:
        rr = tf.sqrt(tf.square(x_channel) + tf.square(y_channel))
        res = tf.concat([res, rr], axis=-1)
    return res


if __name__ == "__main__":
    import numpy as np
    from matplotlib import pyplot as plt
    img = tf.constant(np.random.normal(size=[2, 32, 32, 1]).astype("float32"))
    img_coord = add_coord_channels(img, with_r=True)
    # actual conv can be added using e.g. tf.layers.conv2d(img_coord, ...)
    sess = tf.Session()
    img_c = sess.run(img_coord)[0,:]
    channels = img_c.shape[-1]
    for i in range(channels):
        plt.subplot(channels, 1, i+1).imshow(img_c[:,:,i])
    plt.show()