Skip to content

Latest commit

 

History

History
143 lines (115 loc) · 4.36 KB

INSTALL.md

File metadata and controls

143 lines (115 loc) · 4.36 KB
Raw

Installation Guide

Prerequisites

The main branch works with PyTorch 1.3.1 or higher.

  • Linux (tested on Ubuntu 16.04.4 LTS)
  • Python: 3.7.6
    • 3.6.9 tested
    • 3.7.6 tested
  • PyTorch: 1.4.0
    • 1.3.1 (with CUDA 10.1, torchvision 0.4.2)
    • 1.4.0 (with CUDA 10.1, torchvision 0.5.0)
  • nvcc: 10.1
    • 9.1.85, 10.1.243 compiling and execution tested
  • gcc: 5.4.0
    • 5.4.0
    • 5.5.0
  • Nvidia driver version: 440.59
  • Cudnn version: 7.6.4
  • Pyenv or Anaconda

Python dependencies list in requirements.txt

New Prerequisites

We now support the higher PyTorch and CUDA version and thus can be run on newer devices.

  • Linux (tested on Ubuntu 20.04 LTS)
  • Python: 3.7.16
  • PyTorch: 1.12.0 (with CUDA 11.3)
  • nvcc: 11.3
  • gcc: 9.4.0
  • Nvidia driver version: 510.73
  • GPU: Nvidia RTX 3090

You could change the versio s of torch and torch-vision in requirements.txt. You may encounter some errors during training and inference with new settings. Try the following solutions:

# numpy related error
pip uninstall numpy
pip install numpy

# numba related error
pip uninstall numba
pip install numba

Installation

  • Clone this repo:
git clone -b main --single-branch https://github.com/SysCV/qd-3dt.git
  • Create folders
cd qd-3dt/
mkdir data work_dirs checkpoints
  • You can create a virtual environment by the following:

We use pyenv as our Python version management tool. If you have experienced some installation issue, please refer to Troubleshooting. You may also use Conda for the following installation.

# Add path to bashrc 
echo -e '\nexport PYENV_ROOT="$HOME/.pyenv"\nexport PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.bashrc
echo -e 'if command -v pyenv 1>/dev/null 2>&1; then\n  eval "$(pyenv init -)"\nfi' >> ~/.bashrc

# Install pyenv
curl -L https://raw.githubusercontent.com/pyenv/pyenv-installer/master/bin/pyenv-installer | bash

# Restart a new terminal if "exec $SHELL" doesn't work
exec $SHELL

# Install and activate Python in pyenv
pyenv install 3.7.6
  • Create virtualenv for training, inference and evaluation.
pyenv virtualenv 3.7.6 qd_3dt
pyenv local qd_3dt
pyenv activate qd_3dt
pip install -r requirements.txt
  • Install requirements, create folders and compile binaries for detection under qd_3dt virtualenv.
bash install.sh

nuScenes

Build new virtualenv for nuScenes 3D tracking evaluation.

  • Create virtualenv and install nuScenes toolkit with motmetrics==1.13.0 for AMOTA evaluation.
pyenv virtualenv 3.7.6 Nusc
pyenv activate Nusc
pip install nuscenes-devkit
pip install -r requirements_nusc.txt
pyenv deactivate
cd scripts/
git clone https://github.com/nutonomy/nuscenes-devkit
cd nuscenes-devkit

# Check to the commit for [email protected] evaluation
git checkout e2d8c4b331567dc0bc36271dc21cdef65970eb7e
cd ../../

Waymo

Build the waymo-open-dataset for local evaluation.

  • First follow local compilation to setup waymo-open-dataset toolkit under scripts/waymo_devkit/.

  • NOTE: you need to checkout to commit b2b2fc8f06ed6801aec1ea2d406d559bff08b6b5 instead of remotes/origin/master for successful installation.

  • Compile required evaluation metric functions for both detection and tracking.

cd scripts/waymo_devkit/waymo-od

bazel build waymo_open_dataset/metrics/tools/compute_detection_metrics_main
bazel build waymo_open_dataset/metrics/tools/compute_tracking_metrics_main
bazel build waymo_open_dataset/metrics/tools/create_submission
${QD-3DT_ROOT}
|-- scripts
`-- |-- waymo_devkit
    `-- |-- waymo-od
        |-- eval_waymo_3d.py
        |-- generate_waymo_gt.py
        |-- submission_3d_det.txtpb
        `-- submission_3d_mot.txtpb