[WIP] Pytorch v1.2

[botcs]

Main issue is uint8 / bool.
unit tests are passing, except checkpointer test: had to modify since the expected behavior of nn.DataParallel has changed.

Feel free to add any modifications

[qizhuli]

I was wondering if it would make more sense to dynamically check pytorch version, and choose torch.bool or torch.uint8, so that the code remains backward compatible, and fewer nasty surprises when updating codebase/pytorch? Could be done by replacing torch.uint8 (in places where they are used for indexing) with a simple helper function mask_dtype(), which looks like this:

import torch
from maskrcnn_benchmark.somewhere import get_version

def mask_dtype():
    major, minor = get_version(torch)[:2]
    if major * 100 + minor * 10 >= 120:
        return torch.bool
    else:
        return torch.uint8

and put this helper function in somewhere:

def get_version(module):
    version = getattr(module, '__version__').split('.')
    parsed_version = []
    for v in version:
        try:
            v = int(v)
            parsed_version.append(v)
        except ValueError:
            parsed_version.append(-1)
    return parsed_version

This can also be used to dynamically control the place we call scheduler.step(), whose behaviour changed in pytorch 1.1.

[botcs]

I like the idea, but I think the same amount of effort would be more useful in adding a few convenience features in detectron2.

There are a lot of stuff still here waiting for cleanup, also detectron2