Which pytorch version should be used?

[phil-bergmann]

Which pytorch version should be used?

[timmeinhardt]

Please have a look at our README.md. It specifies the correct PyTorch version necessary to run this code.

[edithzeng]

I got the same error with CUDA 9.0 and torch==0.3.1 on Ubuntu 16.04.6.

[timmeinhardt]

Which error? The error during compilation or loading of the weights? For the latter, could you please describe in more detail what means "unable to load"?

[Midhilesh29]

I have downloaded CUDA 9.0 NVIDIA docker container and installed Pytorch 0.3.1 and tried to load the FRCNN weights that is provided.

import torch
torch.__version__
'0.3.1'

I got the following error.

import torch
a=torch.load("./output/frcnn/res101/mot_2017_train/180k/res101_faster_rcnn_iter_180000.pth")

/usr/local/lib/python3.6/dist-packages/torch/cuda/__init__.py:95: UserWarning: 
    Found GPU0 GeForce RTX 2080 Ti which requires CUDA_VERSION >= 9000 for
     optimal performance and fast startup time, but your PyTorch was compiled
     with CUDA_VERSION 8000. Please install the correct PyTorch binary
     using instructions from http://pytorch.org
    
  warnings.warn(incorrect_binary_warn % (d, name, 9000, CUDA_VERSION))

I have also checked the CUDA version

root@53eea167b2d4:/workspace/midhilesh/tracking_wo_bnw# nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Sep__1_21:08:03_CDT_2017
Cuda compilation tools, release 9.0, V9.0.176

after this warning in the above torch.load() command was running more that 1 minute and when I checked nvidia-smi, the GPU memory is consumed beyond the size of the .pth file.

The above problem (loading weights) does'nt occur while using pytorch 1.3. But the fasterRCNN code written by Jianwei Yang which is used in this code (tracking without bells and whistles) needs pytorch 0.4.0 for running the code. https://github.com/jwyang/faster-rcnn.pytorch/#prerequisites

[timmeinhardt]

This seems to be a problem of your PyTorch installation. Although ncvv might be linked to the correct CUDA version, PyTorch might not be able to find it and compiles with 8.0 instead of 9.0. What is the output of torch.version.cuda?

I found this issue which states that CUDA 9.0 doesnt work well with your GPU.

[Midhilesh29]

torch.version.cuda
'8.0.61'

So, how to change this cuda version?. If my GPU 2080Ti is not compatible with CUDA 9.0, is there any way to run your code?

[timmeinhardt]

You could try running the code only on CPU. And for the PyTorch cuda version, please double check all the CUDA paths (CUDA_HOME, LD_LIBRARY_PATH, PATH, etc.) before installing PyTorch.

[Midhilesh29]

Thank you for helping me!! I will check it.

[timmeinhardt]

Can we close this issue?

[Midhilesh29]

yes sure, I will close it

[nicolaerosia]

@Midhilesh29 I have a RTX 2070, did you found a workaround? I've opened a new issue which is basically a duplicate of this, #35 and there's also #16