SSL-vs-SSL-benchmark

This is the code for benchmark comparing self-supervised and semi-supervised deep classifiers for medical images

Supplementary Materials

Here we provide the Supplementary Materials[Appendix.pdf/] of our benchmark. The Supplement includes following sections to describe the experiments and analysis in more details.

Code and Data Splits for Reproducibility
Dataset Details
Additional Results
Algorithms Details
Additional Analysis

Setup

Prepare datasets

TissueMNIST and PathMNIST: please visit https://zenodo.org/record/6496656
AIROGS: please visit: https://zenodo.org/record/5793241
TMED2: please visit https://TMED.cs.tufts.edu and follow the instruction. We use the split1 in the released data.

Install Anaconda

Follow the instructions here: https://conda.io/projects/conda/en/latest/user-guide/install/index.html

Environment

packages needed are specified in environment.yml (TODO)

Running experiments

Define the environment variable

export ROOT_PATH="paths to this repo" (e.g., '/ab/cd/SSL-vs-SSL-benchmark', then do export ROOT_PATH = '/ab/cd/SSL-vs-SSL-benchmark')

Example

For example if you want to run FixMatch on TissueMNIST to reproduce Figure 1(a) and Figure A.2(a), go to runs/TissueMNIST/FixMatch/ bash launch_experiment.sh run_here

Note that you will need to edit the paths to dataset in the launch_experiment.sh file.

A note on reproducibility

While the focus of our paper is reproducibility, ultimately exact comparison to the results in our paper will be conflated by subtle differences such as the version of Pytorch etc (see https://pytorch.org/docs/stable/notes/randomness.html for more detail).

Acknowledgement

This repository builds upon the public repo pytorch-consistency-regularization[https://github.com/perrying/pytorch-consistency-regularization]. Thanks for sharing the great code bases!

Reference

TODO

tufts-ml/SSL-vs-SSL-benchmark