I can relate. Spent countless hours on this stuff with computer vision and convnets. The intricacies of align_corners, implementation differences between deep learning frameworks, striding and pooling when numbers aren't neatly divisible, uuh.