Upgrade to PyTorch 1.0

[hellock]

This PR solves #155. In fact, this repo can support PyTorch 1.0 without any changes. Only some minor changes are needed for mmcv.

We made some minor updates though.

• Support half tensors (fp16) for roi_align and roi_pooling.
• Remove some legacy code.

You need to use the pytorch-1.0 branch of mmcv with this PR. After the branch being merged, mmcv will be updated to v0.2.3.

The results are shown in the below table.

	PyTorch 0.4.1	PyTorch 1.0.0
RPN	57.1	57.0
Faster R-CNN	36.4	36.5
Mask R-CNN	37.3 / 34.2	37.3 / 34.2
RetinaNet	35.6	35.7
Cascade R-CNN	40.3	40.4
Cascade Mask R-CNN	41.3/35.7	41.2/35.7

[adityaarun1]

@hellock How do one install mmcv>=0.2.2 ? The master branch on its github channel installs version 0.2.1

[hellock]

@adityaarun1 mmcv 0.2.2 is not released yet, you can checkout the pytorch-1.0 branch of mmcv to use it.

[zxt881108]

@hellock How about the training speed after updating pytorch 1.0?

[wangg12]

@hellock mmcv 0.2.2 has been released, is it ready to merge this into master?

[hellock]

Hi @wangg12 , recently released v0.2.2 is the version for PyTorch 1.0. If you'd like to use PyTorch 1.0, you can checkout to the pytorch-1.0 branch of mmcv, which may be released later as v0.2.3.

[hellock]

@zxt881108 Currently the speed of PyTorch 1.0 and 0.4.1 is similar since we made only minor changes, we will take time to optimize it.