search

daniilidis-group / neural_renderer

A PyTorch port of the Neural 3D Mesh Renderer
Other
1.12k stars 248 forks source link

pytorch verssion 0.4.0 and CUDA runtime version #61

Closed songweige closed 5 years ago

songweige commented 5 years ago

Hi! It looks like you've replaced AT_ASSERTwith AT_CHECK in the PyPI version. So I was not able to install through pip with pytorch==0.4.0 on my machine. If that is the case, could you please update the Readme file?

When I switched to pytorch==0.4.1, I could install the package but was not able to run the examples. Here are the errors I got:

Error in forward_face_index_map_1: CUDA driver version is insufficient for CUDA runtime version
Error in forward_face_index_map_2: CUDA driver version is insufficient for CUDA runtime version
Error in forward_texture_sampling: CUDA driver version is insufficient for CUDA runtime version

I'm sure that my pytorch is successfully installed. Any suggestions are highly appreciated!

songweige commented 5 years ago

Here are some other things I've tried that didn't work:

Since I have multiple CUDA on my server, I tried to install with specific CUDA path: CUDA_HOME=/share/apps/cuda-9.0/ pip install neural_renderer_pytorch. But it didn't work.

Tried with CUDA 1.0.1 but got an ImportError: ImportError: /home/songweig/anaconda2/envs/pytorch1.0.1/lib/python3.6/site-packages/neural_renderer/cuda/load_textures.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2at15UndefinedTensor10_si ngletonE

songweige commented 5 years ago

I finally made it work with Python2.7+CUDA9.0.176+Pytorch1.1.0. It looked like torch CUDAextension somehow used unexpected CUDA path if there are multiple CUDA installed.