TEACUP: Training-free Assembly as Clinical Uncertainty Predictor

TEACUP is a fast and accurate framework for automated clinical deep learning that characterized the quality of deep networks without training.

Getting Started

Prerequisites

You will need to download AMBER and switch to the branch specifically designed for our project.

git clone https://github.com/zj-zhang/AMBER.git
cd AMBER
git checkout ntk

Then, please add amber to your Python path.

Installation

1. Clone the repo

git clone git@github.com:zhanglab-aim/TEACUP.git

2. Install conda environments

cd TEACUP
conda create -n teacup conda.yml

3. Install further packages

conda activate teacup
pip install thop medmnist

Data Downloading

For ECG data, please download the file ecg/challenge2017.pkl from this link. The dataset is provided from NAS-Bench-360. Then, make a dat/nas-bench-360 folder to store the file.

mkdir -p data/nas-bench-360
mv [location of challenge2017.pkl] data/nas-bench-360/

For CT data, it will be automated downloaded by medmnist package. No further action needed.

Run TEACUP

Reinforcement Learning

python run_genetic.py --dataset ECG2017 --store results --verbose
python run_genetic.py --dataset NoduleMNIST3D --store results --verbose

Genetic Algorithm

python run_genetic.py --dataset ECG2017 --store results --verbose
python run_genetic.py --dataset NoduleMNIST3D --store results --verbose

Pruning Method

python run_pruning.py --dataset ECG2017 --store results --verbose
python run_pruning.py --dataset NoduleMNIST3D --store results --verbose

Contact

If you encounter any issues and/or would like feedbacks, please leave a GitHub issue. We will try to get back to you as soon as possible.