An error appears during the training that may pass in a non-contiguous input. about improved-body-parts HOT 8 CLOSED

Hi, I'm glad you are interested in this repo. It seems that many reasons could arise this error. For example:
https://discuss.pytorch.org/t/resolved-batchnorm1d-cudnn-status-not-supported/3049
I use Cuda 10.1 and Cuddn 7.4. And Pytorch 1.0, 1.1, 1.2, 1.3 and 1.4 work fine. Please first make sure the packages' versions are proper before other tries.

from improved-body-parts.

Hi :)
I have received you email. I'm using HDF5 1.10 (better multiprocessing handling) that supports SWMR mode.
https://github.com/jialee93/Improved-Body-Parts/blob/316e71fa93e1dc444b1cfd4fc312c21c13bfe93f/py_cocodata_server/py_data_iterator.py#L42

There is a good discussion here,
and I concluded the discussion here.

from improved-body-parts.

Thanks for your response : )
I find that even though I use h5py downloaded by pip (version 2.10.0), it can also supports SWMR mode, is it still necessary to install HDF5 to rebuilt h5py ?
I also test the train_parallel.py, same error occurs T_T.
BTW, I use Cuda 10.1 and Cndnn 10.0, Pytorch 1.4.0

from improved-body-parts.

I have rebuilt the h5py by HDF5, but the error still occurs.
It's hard to locate the problem T_T.

from improved-body-parts.

Sorry for what you are suffering😓. I have never met such errors before. I haven't used Cndnn 10.0. What if you set torch.backends.cudnn.enabled = False?

from improved-body-parts.

I have tried it before, but it seems useless.

from improved-body-parts.

:( Then, I have no idea for now. Feel free to discuss if more information is found.

from improved-body-parts.

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

from improved-body-parts.