SDBH-PSO Ensemble

Quick start

1. Install CUDA

2. Install PyTorch

3. Install dependencies

pip install -r requirements.txt

4. Preparing datasets and pre-training models

(1) If training, validation and testing are done on the same dataset (Kvasir or CVC-ClinicDB is recommended for this dataset), put the dataset into "./data1", and train_with_data1.py will automatically split the dataset into training, validation and testing according to 8:1:1.
(2) If training, validation and testing are not in the same dataset or in the same center, put the dataset for training and validation into "./data2/train_and_val" and the dataset for testing into "./data2/test", and train_with_data2.py will split the dataset for training and validation by itself according to 9:1:1.
(3) The datasets used in this study are publicly available at:
Kvasir-SEG: https://datasets.simula.no/kvasir-seg/.
CVC-ClinicDB: https://polyp.grand-challenge.org/CVCClinicDB/.
ETIS-LaribpolypDB: https://drive.google.com/drive/folders/10QXjxBJqCf7PAXqbDvoceWmZ-qF07tFi?usp=share_link.
CVC-ColonDB: https://drive.google.com/drive/folders/1-gZUo1dgsdcWxSdXV9OAPmtGEbwZMfDY?usp=share_link.
PolypGen: https://www.synapse.org/#!Synapse:syn45200214.\ (4) Pre-training models should be downloaded via their github connection and placed in location "./Models" after downloading.
https://github.com/microsoft/Swin-Transformer
https://github.com/whai362/PVT
https://github.com/zengjixiangnfft/ESFPNet (Not official MixTransformer, but ESFPNet is excellent!) \

5. run training:

python train_with_data1.py --amp -Ename t16K -e 100 -b 4 -n1 TrebleFormer_L -n2 TrebleFormer_S -n3 FCBFormer_L -n4 FCBFormer_S -n5 ESFPNet_L -n6 ESFPNet_S -nN 6
python train_with_data2.py --amp -Ename t16K -e 100 -b 4 -n1 TrebleFormer_L -n2 TrebleFormer_S -n3 FCBFormer_L -n4 FCBFormer_S -n5 ESFPNet_L -n6 ESFPNet_S -nN 6