README.md

Confidence score Weighting Adaptation using the JMDS (CoWA-JMDS)

This repository is the official implementation of "Confidence Score for Source-Free Unsupervised Domain Adaptation", accepted in ICML 2022. This code refers to SHOT implementation.

Environments

Fix 'prefix' in environment.yaml file.

prefix: /home/[your_username]/anaconda3/env/CoWA

Then create the environment.

$ conda env create --file environment.yaml

Datasets

You can download datasets here:

• VISDA-2017

Training

After downloading the datasets, create following files and directories in this directory.

$ mkdir ./data

data
└── VISDA-C
    ├── train_list.txt
    └── validation_list.txt

Each list.txt file has (image_path, class index) pairs

./data/VISDA-C/train_list.txt

line 1 : /home/[username]/data/VisDA-2017/train/aeroplane/src_2_02691156_4def53f149137451b0009f08a96f38a9__44_349_150.png 0
line 2 : /home/[username]/data/VisDA-2017/train/aeroplane/src_1_02691156_5d0d3f54c5d9dd386a1aee7416e39fad__180_236_150.png 0
...

Then run a script file.

$ chmod +x run_visda.sh
$ ./run_visda.sh