install_pytorch.md

Debugging Common issues

You might see some problems with PyTorch and other python code installation. This document gives some advice.

Make sure cuda and torch versions match

Make sure that cuda and pytorch versions are the same in your installation:

~/src/home-robot$ conda list | grep torch
ffmpeg                    4.3                  hf484d3e_0    pytorch
pyg                       2.3.1           py39_torch_1.13.0_cu117    pyg
pytorch                   1.13.1          py3.9_cuda11.7_cudnn8.5.0_0    pytorch
pytorch-cluster           1.6.1           py39_torch_1.13.0_cu117    pyg
pytorch-cuda              11.7                 h778d358_5    pytorch
pytorch-mutex             1.0                        cuda    pytorch
pytorch-scatter           2.1.1           py39_torch_1.13.0_cu117    pyg
pytorch3d                 0.7.4           py39_cu117_pyt1131    pytorch3d
torchaudio                0.13.1               py39_cu117    pytorch
torchvision               0.14.1               py39_cu117    pytorchs

The numbers after cu should be the same, i.e. cu117.

This should be true even after installing ROS noetic, e.g. with robostack:

# Installs ROS via Rboostack
mamba env create -n home-robot -f src/home_robot_hw/environment.yml

Help Installing PyTorch

You might see some problems with installing

Installing Pytorch

See here to install PyTorch. Example command:

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

PyTorch3d

To install PyTorch3d, run:

conda install pytorch3d -c pytorch3d

If this causes trouble, building from source works reliably, but you must make sure you have the correct CUDA version on your workstation:

pip install "git+https://github.com/facebookresearch/pytorch3d.git"

See the PyTorch3d installation page for more information.