Skip to content

env.md

Latest commit

 

History

History
146 lines (104 loc) · 3.68 KB

env.md

File metadata and controls

146 lines (104 loc) · 3.68 KB

Environment Setup (conda + uv)

This repo contains two upstream-derived codebases with different dependency stacks. Keep them in separate environments:

  • Codebase: mmdetection3d/ (MMDet3D). Conda env name: mm3d.
    • Runs PV-RCNN baseline + RefineMoE variants for KITTI in this repo.
  • Codebase: FSHNet/ (OpenPCDet-style). Conda env name: fshnet.
    • Runs VoxelRCNN baseline for KITTI and FSHNet baselines (primarily Waymo).

Principles:

  • Use conda to manage python (and isolate the environment).
  • Install PyTorch (+ CUDA wheel) manually for your machine.
  • Use uv to install Python package requirements.

Prerequisites

  • conda (Miniconda/Mambaforge)
  • CUDA toolkit/driver matching your chosen PyTorch wheels (if using GPU)

mmdetection3d codebase (env: mm3d)

Create the environment:

conda create -n mm3d python=3.8
conda activate mm3d

python -m pip install -U pip uv

Install PyTorch (pick versions that match your CUDA/driver):

# Example (update as needed):
python -m pip install \
  torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 \
  -f https://download.pytorch.org/whl/torch_stable.html

Install OpenMMLab stack (as used by this repo):

python -m pip install -U openmim 'numpy==1.23.0'
mim install mmengine
mim install 'mmcv==2.0.0rc4'
mim install 'mmdet==3.0.0'

Install this repo's mmdetection3d/ codebase:

# Install test deps (includes pytest/yapf/isort/flake8 via requirements)
uv pip install -r mmdetection3d/requirements/tests.txt

# Editable install (use mim so that `mim download/search` can find model-index metadata)
cd mmdetection3d
mim install -e . --no-build-isolation

Quick sanity check:

pytest mmdetection3d/tests/test_models/test_voxel_encoders/test_pillar_encoder.py

Notes:

  • Optional performance deps (e.g. cumm, spconv) depend on CUDA; install them only if your experiment/config requires them.

FSHNet codebase (env: fshnet)

Create the environment:

conda create -n fshnet python=3.8
conda activate fshnet

python -m pip install -U pip uv

Install PyTorch (pick versions that match your CUDA/driver):

# Example (update as needed; repo authors tested torch>=1.10 for this subtree):
python -m pip install \
  torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 \
  -f https://download.pytorch.org/whl/torch_stable.html

Install FSHNet/ dependencies and build extensions (tested recipe):

cd FSHNet

# CUDA extensions (pick the CUDA-tagged wheels that match your system)
uv pip install \
  cumm-cu113 \
  spconv-cu113 \
  scikit-image \
  pyyaml \
  tensorboardX \
  easydict \
  pyquaternion \
  opencv-python \
  protobuf==3.20.3 \
  triton==2.1.0 \
  numba==0.48.0

# Build and install (editable)
python setup.py develop

# Waymo extras (only needed for Waymo experiments)
python -m pip uninstall -y waymo-open-dataset-tf-2-4-0
uv pip install waymo-open-dataset-tf-2-11-0

# SharedArray: install with plain pip (uv install may fail)
python -m pip install SharedArray==3.1.0

# Install torch-scatter (will resolve a compatible build for your torch/cuda)
uv pip install torch-scatter

Quick sanity check:

python -c "import pcdet; print('pcdet ok')"

Waymo-specific extras:

  • Waymo and related packages are only needed for Waymo experiments.
  • Prefer documenting these as an opt-in block in your experiment notes, since they are platform- and CUDA-sensitive.

Optional dependencies:

  • navsim is optional. If you need navsim==2.0.0, it is not available on PyPI and must be installed from source.

Note:

  • FSHNet/ does not ship a solver-friendly requirements.txt in this repo. Use the recipe above (it matches the original working environment).