Awesome

MICLe_Pytorch

Unoffitial PyTorch implementation of MICLe algorithm in the paper Azizi, S., Mustafa, B., Ryan, F., Beaver, Z., Freyberg, J., Deaton, J., ... & Norouzi, M. (2021). Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 3478-3488).

Requirements

• Python 3.7
• Pytorch 1.10.*
• Pytorch-ligntning 1.6.*
• lightly 1.2.*
• Pillow 8.4.*

Usage

1. Prepare the dataset

You can arrange your dataset in the following structure:

├── dataset
│   ├── subject 1
│   │   ├── image 1
│   │   ├── image 2
│   │   ├── ...
│   ├── subject 2
│   │   ├── image 1
│   │   ├── image 2
│   │   ├── ...
│   ├── ...

To train, run the following command with the path to your dataset:

python main.py --train_dir_path /your/path/train_dataset

You can also specify the hyperparameters in the command line.

The model will be saved as a .pth in your project directory.