Is ordinary convolution（with bias=0 and no use BN） and partial convolution the same?

[luowy1001]

Thank you for your genius work. However, there's a little problem that puzzles me.
For image inpainting task, when the pixel values the missing area of the input image are set to 0,
I think that ordinary convolution(bias=False and no use BN) and partial convolution are the same? Under this circumstance, the result of the ordinary convolution(bias=False and no use BN) is 0 naturally, at the missing area (if all piexls in this slide windows are 0) . And the gradients at these missing areas is 0, it is the same as parital convolution. I'm confused about this small question. Can you tell me the essential difference between them? Thanks!
@liuguilin1225

[liuguilin1225]

In that case, the mask updating will matter. Partial conv will track how the mask evolves when you apply a sequence of partial conv layers.
Usually if you apply an ordinary conv layer without bias, you will follow a BN layer; such BN layer will add transformation onto your whole feature map (including non-holes, filled holes and un-filled holes). A corresponding mask will help you identify which regions are still holes and set them to be 0 again before feeding into next layer. But if you don't track the mask, the un-filled holes are now filled with the transformation from BN, which might confuse the network.

[liuguilin1225]

BTW, for the question regarding to no BN, if one only uses conv layers without bias and no BN etc, the network can't learn anything beyond multiplication.