Learning with noisy labels via Mamba and entropy KNN framework

This repository contains the code and model weights for the RelabelMamba architecture, which introduces a Augment Entropy-based K-Nearest Neighbors Relabel (AEKR) method for correcting noisy labels and an Adaptpuzzle data augmentation strategy. This method is built on Vcm, which leverages Convolutional Networks (ConvNet) and Vision Transformer (ViT) blocks for feature extraction and processing.

Installation
Requirements
Model Architecture
Training Testing
File Structure

Installation

To install the necessary environment for running the code, you can clone the repository and install the required dependencies.

git clone https://github.com/albert-jin/RelabelMamba.git
cd RelabelMamba

Install the required Python packages using pip:

pip install -r requirements.txt

Vim Installation

RelabelMamba utilizes Vim as a backbone for its architecture. You can install the official Vim repository using:

git clone https://github.com/hustvl/Vim.git
cd Vim
pip install -e .

For more detailed installation instructions, please refer to the official Vim repository here.

Requirements

Python 3.8+
PyTorch 1.10.0+
NVIDIA CUDA 11.0+ (if using GPU)
Other dependencies are listed in requirements.txt

Model Architecture

The RelabelMamba architecture integrates a Vim backbone with several advanced strategies for handling noisy labels. Below is an illustration of the overall architecture:

The key components of the architecture include:

Vision Convolutional Mamba: These blocks are responsible for extracting hierarchical features from the input images. They use depthwise separable convolutions and State Space Model (VSS) blocks for efficient feature extraction.
Augment Entropy-based K-Nearest Neighbors Relabel (AEKR): This module dynamically corrects noisy labels by computing the entropy variance and applying label correction based on the k-nearest neighbors in the feature space.
Adaptpuzzle Data Augmentation: This module generates mixed examples based on clean and noisy labels to enhance the robustness of the model to label noise.

Training Testing

To train the model on your dataset, you can run the following command:

python cifar 10.py --dataset_path /path/to/cifar 10 \
               --noisy_dataset_path /data/data/noisy \
               --noise_mode sym \
               --noise_ratio 0.3 \
               --k 200 \
               --epochs 50 \
               --lr 0.001 \

Please ensure that your dataset is properly formatted and placed in the data/ directory. You can modify the training configuration in the provided YAML files.

File Structure

The repository is organized as follows:

RectMamba/
│
├── utils/                  # Configuration files for training and testing
├── datasets/                     # Directory for datasets
├── mamba-1p1p1/        # mamba tools 
├── vim/        # Vision mamba package and pretrained weights directory
├── models/                   # Directory containing the RectMamba model architecture
├── wandb/                  # Results directory for saving outputs
├── cifar 10.py                  # Script for training & testing the model
├── requirements.txt          # Required Python dependencies
├── README.md                 # Project readme

## Contributors ✨

Thanks to these amazing contributors for their hard work on this project:

- [Ningwei Wang](https://github.com/ningwei-wang) (major contributions made)
- [Weiqiang Jin](https://github.com/albert-jin) (minor contributions & supports)

Feel free to check out their profiles for more awesome work! 🎉

albert-jin/RelabelMamba