Spotted by InnovArul/Arulkumar on the forums, thank you!

The documentation for avg_pool2d/3d says

count_include_pad: when True, will include the zero-padding in the
        averaging calculation. Default: True

Indeed, the default was just that up to PyTorch 0.2.
In 0.3-0.4 the default is False for 2d/3d and True 1d and the doc says it was True for all. The change in default happened when nn.yaml replaced the previous mechanism of calling THNN.

Now I was about to submit a PR reverting the default to True, but the subtle twist is that TF uses False by default.
avg_pool2d is used in the wild, e.g. by ResNet.

Should I change the docstring or should I change the default in nn.yaml?
Do you have an opinion (@soumith @colesbury maybe)?

Best regards

Thomas