Skip to content

Latest commit

 

History

History
55 lines (37 loc) · 1.56 KB

INSTALL.md

File metadata and controls

55 lines (37 loc) · 1.56 KB
Raw

Installation

The code was tested with Anaconda Python 3.6, CUDA 10.0, and PyTorch v1.3. Check your gcc version by 'gcc -v'. gcc version may need to be higher than v4.8 in order to compile the DCNv2 package. We tested the code with both gcc v5.4.0 and v8.4.0. After installing Anaconda:

  1. [Optional but highly recommended] create a new conda environment.

    conda create --name trades python=3.6

    And activate the environment.

    conda activate trades

  2. Install PyTorch:

    conda install pytorch=1.3.1 torchvision=0.4.2 cudatoolkit=10.0.130 -c pytorch

  3. Install COCOAPI:

    pip install cython; pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

  4. Clone this repo:

    git clone https://github.com/JialianW/TraDeS.git

  5. Install the requirements

    cd $TraDeS_ROOT
pip install -r requirements.txt

  6. Compile deformable convolutional (Successuflly compiled with both gcc v5.4.0 and v8.4.0. gcc version should be higher than v4.8).

    cd $TraDeS_ROOT/src/lib/model/networks/DCNv2
. make.sh

    (modified from DCNv2)

Note: We found the nuScenes and MOT dataset API versions are not compatible, you can switch between them by running 'sh mot_switch_version.sh' (for MOT experiments) or 'sh nuscenes_switch_version.sh' (for nuScenes experiments). The default installed versions are for MOT dataset.