Skip to content

Latest commit

 

History

History
90 lines (62 loc) · 3.95 KB

README.md

File metadata and controls

90 lines (62 loc) · 3.95 KB

Self-Supervised Learning for End-to-End Particle Reconstruction Using Resnet15 and GAN model

Project Overview

This project involves training a ResNet-15 model for end-to-end particle reconstruction in the CMS (Compact Muon Solenoid) experiment using self-supervised learning. The project is divided into two main stages:

  1. Self-Supervised Learning: Train a ResNet-15 model using the Barlow Twins self-supervised learning method on a provided unlabelled dataset.
  2. Finetuning and Evaluation: Fine-tune the self-supervised model on a labeled dataset and compare its performance with a model trained from scratch.

The dataset generation includes synthetic data created using a GAN model fine-tuned on random images from PyTorch libraries. The final dataset consists of 3-channel images.

Files Provided

  • self_supervised_training.ipynb: Jupyter notebook containing the implementation for Barlow Twins training, finetuning, and evaluation.
  • model_weights/: Directory containing the weights for both the self-supervised model and the fine-tuned model.
  • README.md: This README file.

Data Preparation

Unlabelled Dataset (Pretraining Stage)

  • Used for self-supervised learning.
  • Generated using a GAN model fine-tuned on random images from PyTorch's image libraries.
  • Consists of 3-channel images.

Labelled Dataset (Finetuning Stage)

  • Contains labeled images for supervised learning.
  • The data split is 80% for training and 20% for evaluation.
  • Ensure proper preprocessing and normalization.

Self-Supervised Learning with Barlow Twins

Barlow Twins is a self-supervised learning method that focuses on maximizing the similarity between different augmented views of the same image while minimizing the redundancy between features.

Steps:

  1. Data Augmentation:

    • Apply various augmentations to create different views of the same image.

  2. Model Architecture:

    • Train a ResNet-15 model with Barlow Twins objectives, which include:
      • Variance Minimization: Ensuring features have a low variance.
      • Invariance Maximization: Ensuring similar views have similar representations.
      • Covariance Regularization: Reducing redundancy between feature dimensions.

  3. Training:

    • Train the ResNet-15 model on the unlabelled dataset using Barlow Twins objectives.

  4. Save Model:

    • Save the trained model weights for use in the finetuning stage.

Finetuning and Evaluation

  1. Finetuning:

    • Load the self-supervised ResNet-15 model and finetune it on the labeled dataset.
    • Training is performed with a low learning rate to adapt the model to the labeled data.

  2. Comparison:

    • Train a ResNet-15 model from scratch on the labeled dataset for comparison.
    • Both models are evaluated on 20% of the data reserved for testing.

  3. Evaluation Metrics:

    • Accuracy and loss are recorded for both models.
    • Ensure to avoid overfitting on the test dataset by using an independent sample.

Jupyter Notebook Instructions

  1. Setup:

    • Load the provided Jupyter notebook self_supervised_training.ipynb.
    • Ensure all dependencies (e.g., PyTorch, torchvision, Barlow Twins library) are installed.

  2. Run the Notebook:

    • Follow the instructions in the notebook to run the code.
    • The notebook covers:
      • Training the ResNet-15 model with Barlow Twins on the unlabelled dataset.
      • Finetuning the model on the labeled dataset.
      • Comparing performance with a model trained from scratch.

  3. Model Weights:

    • The model_weights/ directory contains:
      • self_supervised_model.pth: Weights for the self-supervised ResNet-15 model.
      • finetuned_model.pth: Weights for the finetuned ResNet-15 model.

Notes

  • Ensure to handle data preprocessing, augmentation, and normalization appropriately.
  • Monitor for overfitting and use an independent sample for validation.

License

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