Pytorch 1.9 and Cuda 11?

[johnnynunez]

Is this compatible with these versions?

[penincillin]

@johnnync13
We haven't tried this setting on our side. But we believe it won't be a big issue due to the good backward compatibility of Pytorch. One thing to pay attention is that you might need to install all the requirements step-by-step (especially the Detectron-2, remember to choose the one matched with your Pytorch and Cuda version), instead of using the provided auto-installing script.

[wzyinparis]

I used Pytorch 1.8 and Cuda 11.1. It's compatible.

[italosalgado14]

@wzyinparis Hi!, Did you modify the install script install_frankmocap.sh to achieve a correct installation? I have especially had problems with Detectron2 :(

[johnnynunez]

I have the same problem @italosalgado14

[italosalgado14]

@johnnync13 i run the program in a simple venv enviorment with my general CUDA 11.3 and Pytorch 1.9 (+ cudnn + toolkit) in my laptop with 3070, i dont use conda. Some problems with CUDA-Detectron2-conda: detectron2 dont recognize CUDA 11 correctly in conda enviortment. Specifically, the problem is to recognize the nvidia-toolkit INSIDE a conda enviorment: i dont have solution for that :(

[italosalgado14]

PD: Not necesary to edit install_frankmocap.sh or some other archive in my venv ecosystem (only skip/reinstall installation for detectron2). The main problem is that you verify well that you have installed correctly Nvidia Driver+NvidiaTollkit (and the correct PATH!)+cudnn + corresponding version of Pytorch(e.g. 1.9) + corresponding version of Detectron2 for these version of Pytorch. Greetings!