Bayes Transition Partial-label Learning

Bayes Transition PLL is a novel algorithm for partial-label learning, inspired by advancements in noisy label learning. This project introduces a Bayesian transition approach to improve robustness and accuracy in handling ambiguous and noisy labels.

Overview

• Objective: Develop a Bayesian transition matrix-based method to enhance partial-label learning.
• Key Contributions:
  • Innovative algorithm combining noisy label learning techniques with partial-label learning.
  • Focus on robust statistical modeling and rigorous mathematical foundations.
  • Implementation validated on benchmark datasets.

Repository Structure

├── datasets/             # Data files for training and evaluation
├── utils/                # Auxiliary algorithms
├── loss.py               # Custom loss functions using Bayesian transitions
├── main.py               # Main training and evaluation script
├── resnet.py             # Backbone ResNet model
├── resnet_bayes.py       # Bayesian ResNet implementation
└── README.md             # Project documentation

Quick Start

1. Clone the repository:

git clone https://github.com/KevinCarpricorn/Transition_Matrix_PLL.git
cd Transition_Matrix_PLL

2. Run training:

python main.py --epochs 50 --batch_size 32 --lr 0.001

Results

The proposed algorithm demonstrates:

• High accuracy and robustness in partial-label learning tasks.
• • Superior performance in handling label noise compared to baseline methods.

License

This project is licensed under the MIT License.