I got stuck in val，can you give me some suggest? about v2v-posenet-pytorch HOT 11 CLOSED

pytorch/pytorch#1355 It's a bug.

from v2v-posenet-pytorch.

I have not encountered problems like this. I used PyTorch-0.4.1 + python-3.6 + cuda(92/90/80) with anaconda under ubuntu 16.04. I guess if it is a problem in your PyTorch version or python environment?

from v2v-posenet-pytorch.

Tested under GTX1070 and TITAN XP.

from v2v-posenet-pytorch.

As PyTorch release page shows, there are some Data Loader / Datasets / Multiprocessing bug fixes in PyTorch v0.3.1 and later releases. Which PyTorch version do you use?

from v2v-posenet-pytorch.

I used PyTorch-0.4.1 + python-3.5 + cuda80 under ubuntu 16.04.Tested under GTX1080 and GTX1080Ti，and would you support NYU and ICVL test future?

from v2v-posenet-pytorch.

I'm not sure what exactly the problem is. But it's may not the bug in repo's code, and here are some debug directions:

1. Environment related:

• disable cudnn for batchnorm(or just use cuda or use cpu) to check, as suggested in repo's README. (related discussions pytorch-pose/issues/33, pytorch-pose-hg-3d/issues/16)

• Or, create a new environment with anaconda to see if we can bypass this problem.

2. Configuration related:

• set batch_size = 1, set num_workers=1 to see what happens.

3. Data related:

• use only one or fewer samples to overfit to see what happens.
• use your own small real(or simulated) dataset to see what happens.

from v2v-posenet-pytorch.

Thanks a lot,I will try it ! And would you support NYU and ICVL test future?

from v2v-posenet-pytorch.

I have no plan right now, but it should not be hard to port it from the author's repo.

from v2v-posenet-pytorch.

That's good! Can you train your model now?

from v2v-posenet-pytorch.

Yes

from v2v-posenet-pytorch.

Thanks a lot！

from v2v-posenet-pytorch.