docs/source/getting_started/installation_virtual_env.md

Installation With Virtual Environment

Note: the repo was tested with Python 3.10+ and PyTorch 2.0+.

Create Enviroment

We recommend setting up a separate virtual environment for Kimodo to avoid dependency conflicts.

Using venv

python -m venv venv
source venv/bin/activate

Using Conda

conda create -n kimodo python=3.10
conda activate kimodo

Install Dependencies

Install PyTorch

First, make sure to install a version of PyTorch that works with your system and CUDA version. We suggest anything over PyTorch 2.0. We strongly suggest using a GPU-capable version of PyTorch to generate motions in a reasonable amount of time.

(Optional) Clone Modified Viser Library

The interactive demo relies on a fork of Viser that implements a timeline interface and more. If you want to have an editable install of this version of Viser (i.e., you expect to modify it), clone and install it within the kimodo directory using:

git clone https://github.com/nv-tlabs/kimodo-viser.git
pip install -e kimodo-viser

Install Kimodo

Next, install Kimodo run this command from the base of repo:

pip install -e .

This results in a single editable install for Kimodo and the MotionCorrection package.

If you plan to use the demo, you can instead run:

pip install -e ".[all]"

This will install our Viser fork (if not already installed in the previous step) and the SOMA body model.

Next, head over to the Quick Start page to test out your installation by generating some motions.