Awesome
Requirements
torch 2.0.1, mmcv, mmdet
Dataset
-
run extract_data.py for data pre-processing.
datasets ├── cpm17 │ ├── extract_data.py │ ├── test │ └── train ├── cpm17_test_files.npy ├── cpm17_train_files.npy ├── kumar │ ├── extract_data.py │ ├── images │ └── labels ├── kumar_test_files.npy ├── kumar_train_files.npy ├── pannuke │ ├── extract_data.py │ ├── Fold 1 │ ├── Fold 2 │ ├── Fold 3 │ ├── Images │ └── Masks ├── pannuke123_test_files.npy ├── pannuke123_train_files.npy ├── pannuke123_val_files.npy ├── pannuke213_test_files.npy ├── pannuke213_train_files.npy ├── pannuke213_val_files.npy ├── pannuke321_test_files.npy ├── pannuke321_train_files.npy └── pannuke321_val_files.npy
Training
-
Train the prompter
cd prompter python main.py --config dpa_pannuke123.py --output_dir dpa_pannuke123 # python main.py --config dpa_pannuke213.py --output_dir dpa_pannuke123 # python main.py --config dpa_pannuke321.py --output_dir dpa_pannuke123 -
Use the trained prompter to generate nuclei prompts for the validation and test sets.
python predict_prompts.py --config dpa_pannuke123.py --resume checkpoint/dpa_pannuke123/best.pth # python predict_prompts.py --config dpa_pannuke213.py --resume checkpoint/dpa_pannuke213/best.pth # python predict_prompts.py --config dpa_pannuke321.py --resume checkpoint/dpa_pannuke321/best.pth -
Download SAM's pre-trained weights into segmentor/pretrained and train the segmentor.
cd segmentor torchrun --nproc_per_node=4 main.py --config pannuke123_b.py --output_dir pannuke123_b # torchrun --nproc_per_node=4 main.py --config pannuke213_b.py --output_dir pannuke213_b # torchrun --nproc_per_node=4 main.py --config pannuke321_b.py --output_dir pannuke321_b
Evaluation
see test.sh
Checkpoints
| Kumar | CPM-17 | PanNuke123 | PanNuke213 | PanNuke321 | |
|---|---|---|---|---|---|
| Prompter | OneDrive | OneDrive | OneDrive | OneDrive | OneDrive |
| Segmentor-B | OneDrive | OneDrive | OneDrive | OneDrive | OneDrive |
| Segmentor-L | OneDrive | OneDrive | OneDrive | OneDrive | OneDrive |
| Segmentor-H | OneDrive | OneDrive | OneDrive | OneDrive | OneDrive |
Contact
If you have any questions or concerns, feel free to report issues or directly contact us (Zhongyi Shui shuizhongyi@westlake.edu.cn).
Citation
If you find this code useful for your research, please cite us using the following BibTeX entry.
@article{shui2023unleashing,
title={Unleashing the Power of Prompt-driven Nucleus Instance Segmentation},
author={Shui, Zhongyi and Zhang, Yunlong and Yao, Kai and Zhu, Chenglu and Sun, Yuxuan and Yang, Lin},
journal={arXiv preprint arXiv:2311.15939},
year={2023}
}