Contrastive Learning with Image Clustering

This project implements a Contrastive Learning architecture using SimCLR in PyTorch. The model learns meaningful embeddings by contrasting positive and negative pairs of images. The learned embeddings are then clustered using K-Means to evaluate their performance in unsupervised image classification.

Features

• Implements SimCLR contrastive learning from scratch.
• Uses ResNet-18 as the backbone for feature extraction.
• Evaluates embeddings using K-Means clustering and computes metrics such as:
  • Normalized Mutual Information (NMI)
  • Adjusted Rand Index (ARI)
• Visualizes clustering performance using t-SNE for dimensionality reduction.

Dataset

• CIFAR-10 is used as the dataset, containing 60,000 images in 10 categories.
• Images are preprocessed with data augmentations during training and normalized during evaluation.

Results

• The model achieves clear separation of clusters, as shown in the t-SNE plot (refer to clusters_visualization.png).
• Metrics:
  • NMI: ~0.7 (depends on training setup)
  • ARI: ~0.6 (depends on training setup)

How to Run the Code

1. Clone the repository:

git clone https://github.com/your-username/contrastive-learning-clustering.git
cd contrastive-learning-clustering

2. Install dependencies:

pip install -r requirements.txt

3. Train the SimCLR model:

python train_simclr.py

4. Cluster and evaluate embeddings:

python contrastive_learning_with_clustering_and_visualization.py

5. View Clustering Results:

• Check clusters_visualization.png for the t-SNE visualization.

Files in the Repository

• train_simclr.py: Script to train the SimCLR model.
• contrastive_learning_with_clustering_and_visualization.py: Evaluates embeddings using K-Means and visualizes clusters.
• requirements.txt: Python dependencies.
• README.md: Documentation for the project.

Future Improvements

• Use advanced clustering methods (e.g., DBSCAN, hierarchical clustering).
• Test on larger datasets (e.g., ImageNet) for better generalization.
• Extend visualization by overlaying image thumbnails.

License

This project is licensed under the MIT License.