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

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.
        
        Aurther/Source: N-McA, https://gist.github.com/N-McA/9bd3a81d3062340e4affaaaaad332107
        """
        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)
        batch_size = dynamic_input_shape[0]
        w = dynamic_input_shape[-3]
        h = dynamic_input_shape[-2]
        bias = _kb_grid_coords(width=w, height=h)        
        
        bias = kb.expand_dims(bias, 0)
        bias = kb.tile(bias, [batch_size,1,1,1])

        return kb.concatenate([x, bias], axis=-1)

    def compute_output_shape(self, input_shape):
        batch_size, w, h, channels = input_shape
        return (batch_size, w, h, channels + 2)