Comment on CR for #20105:
@gchanan:
testing? It also seems like you don't support a larger number of cases, e.g. depthwise, transposed, dilated convolutions.Maybe similar on the AvgPool / SumPool too, but I'm not sure.
@gchanan:
I don't think this makes sense to implement AvgPool for long because of the integer division (i.e. we don't support long on mean). Doesn't it make sense to implement SumPool instead?

About @gchanan comment on AvgPool for Long: I think it should be implemented. Bilinear / bicubic interpolation on integer types (uint8 for example) are widely used, and I don't think that AvgPool is that different from a bilinear downsampling.

@fmassa do any of those have division on integer dtypes? From what I can tell, those only have division in backwards, and we don't support backwards on integer dtypes anyway.

My naive interpretation would be AvgPool would actually return a non-integer dtype on integer dtype inputs, but we don't have (many) cross-dtype operations yet.

Thoughts?

@gchanan interpolate multiplies the values of the to be interpolated by a floating point value, so this is the same situation as in here. Note that our interpolate kernels do not support integer types yet, but I think they should. Interpolation for integer types is implemented in opencv and Pillow through, and I can grab some examples once I'm back to my computer

But you do bring valid points though. Im not sure if we should stick with what is currently done in image processing libs, or adopt a more numpy-correct alternative

@pytorchbot retest this please