/DACS

Code from the paper "DACS: Domain Adaptation via Cross-domain Mixed Sampling"

Primary LanguagePythonMIT LicenseMIT

Code used for the results in the paper "DACS: Domain Adaptation via Cross-domain Mixed Sampling"

Getting started

Prerequisite

  • CUDA/CUDNN
  • Python3
  • Packages found in requirements.txt

Run training and testing

Example of training a model with unsupervised domain adaptation on GTA5->CityScapes on a single gpu

python3 trainUDA.py --config ./configs/configUDA.json --name UDA

Example of testing a model with domain adaptation with CityScapes as target domain

python3 evaluateUDA.py --model-path checkpoint.pth

Pretrained models

Pretrained model for GTA5->Cityscapes can be downloaded at: (Link), and should be unzipped in the '../saved' folder. This model peaked at 53.66 mIoU and ended at 53.04 mIoU.