Error in Mask Implementation

[loeweX]

There is an error in the implementation of the masks on the horizontal and vertical stacks:

If I use the code to recreate the results from here, I get the following output:

As you can see, there is no information propagated along the first horizontal line. This can be solved by exchanging the code in https://github.com/anantzoid/Conditional-PixelCNN-decoder/blob/055dab695f5754e0787a4a22899bc4918e3d2e1b/models.py#L33-L40 by

with tf.variable_scope("v_stack_1"+i): 
    v_stack_1 = GatedCNN([1, 1, conf.f_map], v_stack_in, False, gated=False, mask=None).output()

with tf.variable_scope("h_stack"+i):
    h_stack = GatedCNN([filter_size if full_horizontal else 1, filter_size, conf.f_map], h_stack_in, True, payload=v_stack_1, mask=mask).output()

with tf.variable_scope("h_stack_1"+i):
    h_stack_1 = GatedCNN([1, 1, conf.f_map], h_stack, True, gated=False, mask=None).output()

i.e. not applying the mask for convolutions with kernels of size 1x1. With this implementation I get the desired output:

[anantzoid]

Thanks for pointing this out @loeweX. Surprisingly, I still got some results using the former implementation. Did you see any improvement using your method?

[loeweX]

I was using a different dataset, so I can't say anything about the numerical results. However, I first noticed the mistake because errors kept propagating along the horizontal lines of images. But this is really only visible when you use very simple images with prolonged edges and might therefore not play a big role when evaluating datasets like Cifar10.

Also, it increases the runtime a lot, as it basically doubles the size of the network to be evaluated.

[anantzoid]

Thanks for clarifying. Can you send a pull request?