The development of computer vision solutions for gigapixel images in digital pathology is hampered by significant computational limitations due to the large size of whole slide images. An appealing solution is to use smaller resolutions during inference, and using knowledge distillation to enhance the model performance from large to reduced image resolutions. We propose to distill this information by incorporating attention maps during training, using strictly positive gradients, via the propsoed AT+ term.
You can find more details in the manuscript: | Conference | ArXiv |
- Install in your enviroment a compatible torch version with your GPU. For example:
conda create -n kd_env python=3.8 -y
conda activate kd_env
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
- Clone and install requirements.
git clone https://github.com/cvblab/kd_resolution.git
cd kd_resolution
pip install -r requirements.txt
We propose using SICAPv2 for validating the proposed setting, in the challenging task of patch-level grading.
Please, download and prepare the following dataset as indicated:
- Download the datasets.
- Set the dataset in the corresponding folder, at
./local_data/datasets/SICAPv2/.
The proposed setting is largely general, and we encourage you to apply it to any patch-level classification task!
- Teacher model training at the largest image resolution available (i.e. 10x).
python main.py --epochs 20 --input_shape 512 --target_shape 512 --experiment 512_KD_0_FM_0_SR_0_AM_0
- Student model training at smaller resolutions (e.g. 2.5x) - Please, note that the relative weight of FM and AT terms should be empirically fixed through optimization in the validation set.
python main.py --epochs 20 --input_shape 512 --target_shape 128 --experiment 128_KD_0_FM_01_SR_0_AM_10 --experiment_name_teacher 512_KD_0_FM_0_SR_0_AM_0 --alpha_fm 0.1 --alpha_La 10.0
If you find this repository useful, please consider citing this paper:
@article{InterResDistillation2023,
title={Attention to Detail: Inter-Resolution Knowledge Distillation},
author={Rocío del Amor and Julio Silva-Rodriguez and Adrián Colomer and Valery Naranjo},
journal={European Signal Processing Conference (EUSIPCO)},
year={2023}
}
- Code is released under MIT License