Workaround for inplace max pooling issue [merged]

[thatguymike]

CuDNN assumes layers are not being modified in place, thus breaking our index tracking for updates in some cases in Caffe. Until there is a workaround in Caffe (index management) or cuDNN, use Caffe
engine for max pooling, or don't use in place layers after max pooling layers

[shelhamer]

This solution forces all max pooling to be done by the Caffe engine without offering an override. Correctness is the best default, so I'm fine with that, but I wanted to check that this global switch is desired.

Note this is a particular instance of #2015 highlighted in #2688.

[thatguymike]

For now, I think this is the best solution until we understand the way we want to fundamentally handle in-place more generally. There is a performance difference, but for most folks doing training, shouldn't matter much. The fundamental issue with inplace operations is going to bite us again and again I fear. We can choose to try to address the fundamental issue, but without this change or similar workaround we are breaking people's nets.

[shelhamer]

@thatguymike agreed. Please adjust the indentation (4 spaces to align with code) for merge -- for some reason lint is not checking the style.

Thanks for the workaround.

[thatguymike]

The whole file appears to be 2 space indent for code, which is what I have... Or do you just mean you want the comment block indented? Lint is enforcing style, it made me add the other style changes you prefer like spaces after if's.

[shelhamer]

Ok, merged in 37066eb (with changes squashed into a single commit). Closing this PR.

@thatguymike It was the comments. I thought our lint check caught out-of-line comments but apparently not.