Unable to generate results with a model

[ck-amrahd]

Hi guys,
I am now trying to replicate the results on RTX 2080Ti. I added -gencode arch=compute_75,code=sm_75 to the nvcc command to make it work on RTX 2080Ti. I did the bash install and downloaded the models. It goes fine upto this step. Then when I run: bash ./scripts/figures.sh 0 car df, then I get the following error:

ImportError: /home/user/Downloads/VON/render_module/vtn/vtn/_ext/vtn_lib/_vtn_lib.so: undefined symbol: __cudaRegisterFatBinaryEnd

I see that the _vtn_lib.so is created there. Could you please point me in the right direction? Thank you.

[ztzhang]

I think this sounds like your cuda & pytorch & gcc versions are in a weird combination. Does the dock image work for you?

[ck-amrahd]

Hi ztzhang,
Thank you for the reply. The above method didn't work [yes exactly due to the difference in PyTorch and Cuda versions] and I tried with the given Docker image. I had to change it in multiple places to make it work and the installation goes perfectly. But once I run the bash ./scripts/figures.sh 0 car df, I get cudnn_status_execution_failed. I think the problem comes from the RTX 2080Ti graphics cards I have. I will run again with -gencode arch=compute_75, code=sm_75 [I think that's the correct way, please correct me if I am wrong] to compile for the RTX series. I also tried running it with a lower batch size but it didn't work. Thank you.