Concatenates the (x, y) coordinate normalised to 0-1 to each spatial location in the image. Allows a network to learn spatial bias. Has been explored in at least one paper, "An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution" https://arxiv.org/abs/1807.03247

keras_spatial_bias.py

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

def _kb_linspace(num):
    num = kb.cast(num, kb.floatx())
    return kb.arange(0, num, dtype=kb.floatx()) / (num - 1)

def _kb_grid_coords(width, height):
    w, h = width, height
        
    coords = kb.stack(
        [
            kb.reshape(kb.tile(kb.expand_dims(_kb_linspace(num=w), -1), [1, h]), [-1]),
            kb.tile(_kb_linspace(num=h), [w]),
        ],
        axis=-1,
    )
    coords = kb.reshape(coords, [w, h, 2])
    return coords

class ConcatSpatialCoordinate(Layer):

    def __init__(self, **kwargs):
        """Concatenates the (x, y) coordinate normalised to 0-1 to each spatial location in the image.
        Allows a network to learn spatial bias. Has been explored in at least one paper,
        "An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution"
        https://arxiv.org/abs/1807.03247
        
        Improves performance where spatial bias is appropriate.
        
        Works with dynamic shapes.
        
        # Example
        ```python
        x_input = Input([None, None, 1])
        x = ConcatSpatialCoordinate()(x_input)

        model = Model(x_input, x)

        output = model.predict(np.zeros([1, 3, 3, 1]))
        spatial_features = output[0, :, :, -2:]
        assert np.all(spatial_features[0,   0] == [0, 0])
        assert np.all(spatial_features[-1, -1] == [1, 1])
        assert np.all(spatial_features[0,  -1] == [0, 1])

        # Because this example was 3x3, cordinates are [0, 0.5, 1], so
        assert np.all(spatial_features[1,  1] == [0.5, 0.5])
        ```
        """

        
        if kb.image_data_format() != 'channels_last':
            raise Exception((
                "Only compatible with" 
                " kb.image_data_format() == 'channels_last'"))
    
        super(ConcatSpatialCoordinate, self).__init__(**kwargs)

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

    def call(self, x):
        dynamic_input_shape = kb.shape(x)
        w = dynamic_input_shape[-3]
        h = dynamic_input_shape[-2]
        bias = _kb_grid_coords(width=w, height=h)
        return kb.concatenate([x, kb.expand_dims(bias, 0)], axis=-1)

    def compute_output_shape(self, input_shape):
        batch_size, w, h, channels = input_shape
        return (batch_size, w, h, channels + 2)
    
    
def test_ConcatSpatialCoordinate():
    import numpy as np
    from keras.layers import Input
    from keras.models import Model

    x_input = Input([None, None, 1])
    x = ConcatSpatialCoordinate()(x_input)

    model = Model(x_input, x)

    output = model.predict(np.zeros([1, 3, 3, 1]))
    spatial_features = output[0, :, :, -2:]
    
    # The following are always true.
    assert np.all(spatial_features[0,   0] == [0, 0])
    assert np.all(spatial_features[-1, -1] == [1, 1])
    assert np.all(spatial_features[0,  -1] == [0, 1])

    # Because this example was 3x3, cordinates are [0, 0.5, 1], so
    assert np.all(spatial_features[1,  1] == [0.5, 0.5])
    
    
if __name__ == '__main__':
    test_ConcatSpatialCoordinate()

Thanks for sharing!

For visitors that stopped by this page, below fork contains a minor update to method call to account for the batch size while tiling the coordinate layers to the original input. For example, if the input is 16x128x128x1, the output from ConcatSpatialCoordinate would be 16x128x128x3.
https://gist.github.com/pangyuteng/8f4f7c09b490e1baaef852d07105db77