Environment Setup

conda create -n swinunet python=3.8 && conda activate swinunet
pip install -r requirements.txt

Download code from GitHub - HuCaoFighting/Swin-Unet

Pretrained Model & Dataset

Download swin_tiny_patch4_window7_224.pth to pretrained_ckpt/. Download project_TransUNet.zip, extract data/ to root. Rename lists/list_Synapse to lists/Synapse, create val.txt with case IDs.

Code Modifications

• train.py: adjust --num_classes, --n_class, --num_workers, --batch_size
• trainer.py: set num_workers=0

Training

python train.py --max_epochs 20 --base_lr 0.05 --batch_size 12

Models saved to output/ as best_model.pth and last_model.pth.