🧼🔎 SelfClean

A holistic self-supervised data cleaning strategy to detect irrelevant samples, near duplicates, and label errors.

Publications: SelfClean Paper (NeurIPS24) | Data Cleaning Protocol Paper (ML4H23@NeurIPS)

NOTE: Make sure to have git-lfs installed before pulling the repository to ensure the pre-trained models are pulled correctly (git-lfs install instructions).

This project is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International license.

Installation

Install SelfClean via PyPI:

# upgrade pip to its latest version
pip install -U pip

# install selfclean
pip install selfclean

# Alternatively, use explicit python version (XX)
python3.XX -m pip install selfclean

Getting Started

You can run SelfClean in a few lines of code:

from selfclean import SelfClean

selfclean = SelfClean(
    # displays the top-7 images from each error type
    # per default this option is disabled
    plot_top_N=7, 
)

# run on pytorch dataset
issues = selfclean.run_on_dataset(
    dataset=copy.copy(dataset),
)
# run on image folder
issues = selfclean.run_on_image_folder(
    input_path="path/to/images",
)

# get the data quality issue rankings
df_near_duplicates = issues.get_issues("near_duplicates", return_as_df=True)
df_irrelevants = issues.get_issues("irrelevants", return_as_df=True)
df_label_errors = issues.get_issues("label_errors", return_as_df=True)

Examples: In examples/, we've provided some example notebooks in which you will learn how to analyze and clean datasets using SelfClean. These examples analyze different benchmark datasets such as:

• Imagenette 🖼️ (Open in NBViewer | GitHub | Colab)
• Oxford-IIIT Pet 🐶 (Open in NBViewer | GitHub | Colab)

Also, check out our Kaggle notebook to see an illustration of how to get a gold medal for cleaning a competition dataset.

Development Environment

Run make for a list of possible targets.

Run these commands to install the requirements for the development environment:

make init
make install

To run linters on all files:

pre-commit run --all-files

We use the following packages for code and test conventions:

• black for code style
• isort for import sorting
• pytest for running tests