Environment Setup

InfiniCube uses a dual-environment strategy to resolve dependency conflicts between waymo-open-dataset-tf-2-11-0 (requires TensorFlow) and the main project dependencies (PyTorch ecosystem).

Architecture

• Main Environment (conda infini environment): PyTorch, PyTorch3D, gsplat, and other core dependencies
• Waymo Environment (.venv-waymo): Only waymo-open-dataset and its dependencies (TensorFlow, etc.)

---

Setup

Prerequisites

conda env create -f environment.yaml

Create Main Environment

conda activate infini # make sure in the conda environment
uv pip install -e .[main]

# Install pytorch3d from source. takes ~2min
uv pip install --no-build-isolation "git+https://github.com/facebookresearch/pytorch3d.git"
# Install mmcv using mim, then ensure numpy<2.0.0 is respected
mim install "mmcv>=2.0.0"
uv pip install "mmsegmentation>=1.0.0" "numpy<2.0.0"

Create Waymo Environment

This is for data processing. Since waymo-open-dataset has strict version requirements, we create a separate environment to make things easier.

conda activate infini # make sure in the conda environment
uv venv .venv-waymo --python $(which python) # use the same python interpreter as the main environment
uv pip install -e .[waymo] --python .venv-waymo/bin/python # install to the virtual environment

---

Usage

Activate Main Environment

conda activate infini

Use Waymo Environment

You will only need to activate the waymo environment when you are processing waymo data.

conda activate infini && source .venv-waymo/bin/activate

Deactivate the waymo environment (the virtual environment created by uv) when you are done.

deactivate

Install Other Packages

You should use uv pip install to install other packages. It is not recommended to use conda install for python packages now because we are using uv to parse dependencies!