Follow the setup steps on the GitHub homepage of PyTorch.

What I got: Anaconda3, cuda 11.3, cudnn 8.2

check cuda version: nvcc check GPU: nvidia-smi

Install on Windows

conda install astunparse numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing_extensions future six requests dataclasses
conda install -c conda-forge libuv=1.39

Conda did not find I eventually used pip because conda already includes pip. Also, I realized I don’t have to build from source. Using binary installation is much easier.

pip3 install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f

Then I created a new python project for testing installation and it worked. I change the python into anacoda python (shows at the buttom-left of VScode) by modifying python.pythonPath and Ctrl+Shift+P select interpreter. Another setting is switching powershell in Terminal into Cmd prompt of anaconda. Therefore I can use python and venv of anaconda to develop PyTorch.

It taked quite some time to organize the environment. Perhaps I should just use pip and my original python. However, everything seemed ready. Next steps:

  • easy PyTorch examples: play with datasets (e.g., MNIST), learn from examples will be quick
  • we will also need to setup RL env (seems easier?) and play with RL examples.
  • think about the VR project with computational techniques. (reading paper and collect ideas)