ndimage.convolve should not use default mode 'reflect'

[blownhither]

Problem

Gaze-Shift-Net/operations_np.py

Line 106 in a727f63

im1 = ndimage.convolve(im, np.expand_dims(np.stack([tf1]*3, axis=-1), axis=0))

With the stacking, convolution happens between an image of [B, H, W, 3] and kernels of [1, 3, 3, 3]. Note that both operands have 3 channels. And the default mode of ndimage.convolve is 'reflect'. Reflect is weird as detailed in this link, and would behave like this when both operands have 3 elements on some dimension (use 1d array for illustration):

ndimage.convolve([1, 2, 3], [1, 1, 1]) == array([4, 6, 8])
# if 3 channels are convoluted against 3 channels
ndimage.convolve([R, G, B], [1, 1, 1]) == array([R+R+G, R+G+B, G+B+B])  
# [B, G, R | R, G, B | B, G, R]
#      ^ [1, 1, 1]

This is not doing edge detection consistently over all channels, or over each channel. It is a weird combination that involves things like [R+R+G].

Proposal

If I understand it correctly, it would make more sense if we adopt one of these:

• use per-channel edge detection

im1 = ndimage.convolve(im, tf1[:, :, :, np.newaxis])      # kernel [1, 3, 3, 1], no wrapping

• use all-channel edge detection

im1 = ndimage.convolve(im, np.expand_dims(np.stack([tf1]*3, axis=-1), axis=0), mode='wrap')     # wrap as [r,b,g,r,g,b] seqeuence

• edge detection on luminance

More

This seems to be happening for TF operations as well. May change padding='SAME' to 'VALID'. It is very unlikely that this would affect training though.

Gaze-Shift-Net/operations.py

Line 308 in a727f63

im1 = tf.nn.conv2d(im, tf1, strides=[1, 1, 1, 1], padding='SAME')