AWDA

Official PyTorch implementation of AWDA: Adversarial and Weighted Domain Adaptation for Cross-Dataset Change Detection.

This public version contains the training code for the proposed AWDA method with a ResNet-50 backbone.

Installation

conda create -n awda python=3.10 -y
conda activate awda
pip install -r requirements.txt
bash models/backbones/get_resnet50_pretrained_model.sh

The pretrained file is intentionally excluded from Git because it exceeds GitHub's regular file-size limit.

Data layout

Each dataset directory must contain paired images, binary labels, and split files:

DATASET/
|-- A/
|-- B/
|-- label/
`-- list/
    |-- train.txt
    `-- val.txt

Each line in a split file contains three paths relative to the dataset root:

A/0001.png B/0001.png label/0001.png

Update the three data_dir values in configs/awda_resnet50.json before training.

Training

python train.py --config configs/awda_resnet50.json --device cuda