Exploring Label Efficiency with Semi-Supervision and Self-Supervision Methods

This repository is a software framework designed to leverage both Semi-Supervised and Self-Supervised Learning techniques to utilize unlabeled data effectively during the training process. These methods help improve model performance by generating pseudo-labels or creating artificial labels from the data itself, enabling the model to learn useful representations. The framework is built to support a wide range of applications and tasks, providing detailed documentation for projects that aim to incorporate these advanced learning techniques.

Table Of Contents

• License
• Installation
• Getting Started
• Usage
• Project Structure
• Documentation
• Contributing
• Acknowledgments

License

This project is licensed under the MIT License - see the LICENSE file for details.

Please ACKNOWLEDGE THE AUTHOR if you use this repository in your project by including a link to this repository.

Installation

The prerequisites for this project are:

• Python 3.11+
• pip

To install this project, first clone the repository:

git clone https://github.com/xico2001pt/exploring-label-efficiency

Then, install the dependencies:

pip install -r requirements.txt

Getting Started

To adapt this project to your needs, it is recommended to read the documentation file. This file contains a brief overview of the available documentation and links to the different sections of the documentation.

Usage

There are 4 tool scripts in this repository, which correspond to three different learning paradigms and a test script:

• sl_train
• semisl_train
• selfsl_train
• test

To train or test a model, run the following command, where {TOOL_SCRIPT} is one of the tools above and {CONFIG_PATH} is the path for the configuration file:

python -m src.tools.{TOOL_SCRIPT} --config experiments/{CONFIG_PATH}

For example, to use Supervised Learning to train the CIFAR-10 dataset, the following command can be used:

python -m src.tools.sl_train --config experiments/sl/cifar10/wideresnet/sl_cifar10_wideresnet.yaml

Project Structure

exploring-label-efficiency/
├── configs/  # holds the configuration files
│   ├── configs/  # configuration files for the experiments
│   ├── datasets.yaml
│   ├── losses.yaml
│   ├── metrics.yaml
│   ├── models.yaml
│   ├── optimizers.yaml
│   ├── schedulers.yaml
│   ├── selfsl_methods.yaml
│   ├── semisl_methods.yaml
│   └── stop_conditions.yaml
├── data/  # default directory for storing input data
├── docs/  # documentation files
├── logs/  # default directory for storing logs
├── src/  # source code
│   ├── core/  # contains the core functionalities
│   ├── datasets/  # contains the datasets
│   ├── methods/  # contains the SemiSL and SelfSL methods
│   ├── models/  # contains the models
│   ├── tools/  # scripts for training, testing, etc.
│   │   ├── selfsl_train.py
│   │   ├── semisl_train.py
│   │   ├── sl_train.py
│   │   └── test.py
│   ├── trainers/  # contains the trainer classes
│   └── utils/  # utility functions
├── weights/  # default directory for storing model weights
└── requirements.txt  # project dependencies

Documentation

Read the documentation for more details about the project and all the sections mentioned above.

Contributing

If you want to contribute to this project, please contact the author or create a pull request with a description of the feature or bug fix.

Acknowledgments

This repository was developed by Francisco Cerqueira and used this template.