Running with hierarchical blocking

[kentaroy47]

Nice repo, and I enjoyed reading your paper too.

I was wondering how odd-sized blocking can be handled in this repo, since if the feature size cannot be divided by the blocksize it will not use blocking.

def _type_H_block_size(self, h, w):
        return ((int)(h / self.block_size[0]) if h % self.block_size[0] == 0 else h,
                            (int)(w / self.block_size[1]) if w % self.block_size[1] == 0 else w)

This may become troublesome in the last block of resnet, since the feature size would be 512x7x7