Fix tensor type checking for pytorch 1.2

[liyz15]

Fix #725.
In pytorch 1.2, comparison operations return torch.bool instead of torch.uint8, which causes the error.
Ideally we should change all mask to torch.bool. Indexing with torch.uint8 also leads to infinite warning.

[zimenglan-sysu-512]

nice！

[zimenglan-sysu-512]

if u try to replace torch.uint8 with torch.bool, u should be aware of the evaluation in coco_eval.py, in which it needs to set like this:
rles = [
mask_util.encode(np.array(mask[0, :, :, np.newaxis], dtype=np.uint8, order="F"))[0]
for mask in masks
]
for more information, please see this.

[koenvandesande]

I ran into this one as well, ended up with the same fix. Please merge :)

[botcs]

Hi,

Thanks @liyz15 for the fix, I think many of us had this locally, but were too lazy to do a PR :)) so I really appreciate the effort.

@zimenglan-sysu-512 Thank you for pointing this out, I have spent a few hours trying to find out which part I messed up by converting all uint8 to bool.

Contacted @fmassa about PRs waiting for review, hopefully we could move the whole project to use pytorch 1.2 and rewrite the INSTALL.md accordingly soon.

[botcs]

Hi,

I've got permission to merge PRs, so we can start updating the lib to use PyTorch v1.2 (or 1.3?)
There are multiple spots in the source where torch.uint8 is used, and it would be definitely necessary to check if the MODEL_ZOO.md results can be reproduced precisely with higher PyTorch versions.

[ethanweber]

Thanks for the fix!