Skip to content

Latest commit

 

History

History
55 lines (45 loc) · 1.88 KB

INSTALL.md

File metadata and controls

55 lines (45 loc) · 1.88 KB

Installation Guideline

Prerequisites

  • Linux (Ubuntu)
  • Python: 3.8.13
  • PyTorch: 1.11 (with CUDA 11.3, torchvision 0.12.0)
  • PyTorch Lightning: 1.6.5

Installation

  • Install the prerequisites
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt
  • Prepare these directories under the disk with large storage for dataset, checkpoints and visualization.
mkdir ${ASSET_DIR}
mkdir ${WORK_DIR}

cd ${ASSET_DIR}
mkdir data

  • Our code will use asset_dirs and work_dirs under the repo root by default as ${ASSET_DIR} and ${WORK_DIR}, so you might need to symlink them correctly or change the behavior with the config.

  • The experimental results will be saved under ${WORK_DIR}/${EXP_NAME}/${VERSION}. You need to specify the experiment name ${EXP_NAME} and we will use the timestamp as the version name ${VERSION} if you do not config it for every experiment.

Inplace ABN

  • We use Inplace ABN for most of our experiments.

  • Please install it with the latest version.

git clone https://github.com/mapillary/inplace_abn.git
cd inplace_abn
python setup.py install

Seg Fix

  • We use the offset provided from Seg Fix to do the post-processing for our Cityscapes final results.

  • Download the offset_semantic.zip file. Unzip, and place (or symlink) the data as below.

${ASSET_DIR}
└── data
    └── Cityscapes
        ├── leftImg8bit_trainvaltest
        ├── gtFine_trainvaltest
        ├── leftImg8bit_trainextra
        ├── gtCoarse
        ├── refinement
        └── offset_semantic
            ├── val
            └── test_offset