a very very very strange problem on windows about yolact HOT 6 CLOSED

Hmm that is pretty weird indeed. I do most of my testing on Win10 so it's probably not solely a Windows issue. Not to mention, importing Yolact should not instantiate any tensors, so changing the default tensor type before importing should have no effect (key word I guess there is should). Maybe it has to do with your environment?

What's your Pytorch, CUDA, and Python versions? Are you on the latest commit? And when you say "run the eval.py says cuda unkown error, the error locates at...", do you mean line 975 in eval.py?

from yolact.

@dbolya

• try both pytorch cu90-1.0.1 and cu90-1.1.0. by the way, I would like to try cu100-1.1.0,but failed when install...
• python version is 3.6.
• clone the project today...
• run the eval.py in two ways
  1. use cmd, like

     python eval.py --video 'test.mp4' ......

  2. use pycharm, edit the default args' value, then right click,then run..

from yolact.

haha... try another way, edit the yolact.py as follows to make jit disable, then works.

use_jit = torch.cuda.device_count() <= 1
if not use_jit:
    print('Multiple GPUs detected! Turning off JIT.')
use_jit = False
ScriptModuleWrapper = torch.jit.ScriptModule if use_jit else nn.Module
script_method_wrapper = torch.jit.script_method if use_jit else lambda fn, _rcn=None: fn

from yolact.

Okay it looks like this is a known Pytorch issue (pytorch/pytorch#17108) on windows with some Nvidia drivers for Pytorch 1.0.1 and above.

I think the problem is actually with that piece of code there, which would be affected by the above issue. Can you revert your changes and pull my latest commit to see if that workaround fixed this?

from yolact.

@dbolya it works...hahah you're pretty good... thanks a lot.. xiexie dashen

from yolact.

Good to hear! Yeah Pytorch doesn't test as much as it potentially should on Windows.

from yolact.