Indian Classical Dance Classification

This project implements a deep learning model to classify different forms of Indian classical dance using computer vision and neural networks.

Project Overview

The project uses a Convolutional Neural Network (CNN) based on ResNet18 architecture to classify eight different forms of Indian classical dance:

• Bharatanatyam
• Kathak
• Kathakali
• Kuchipudi
• Manipuri
• Mohiniyattam
• Odissi
• Sattriya

Project Structure

indiandanceclassification/
├── CNN.ipynb                              # Main training notebook
├── dance_model.pth                        # Trained model weights
├── gradcam_*.jpg                          # Grad-CAM visualizations
├── resnet18-gradcam-visualization-flat.py # Visualization script
├── README.md                              # Project documentation
└── final/                                 # Dataset directory 
    ├── dataset.py                         # Dataset handling code
    ├── test.csv                          # Test set metadata
    ├── train.csv                         # Training set metadata
    ├── csv2/                             # Additional metadata
    ├── test/                             # Test images
    ├── train/                            # Training images
    └── validation/                       # Validation images

Technical Details

Model Architecture

• Base model: ResNet18 (pretrained)
• Modified final fully connected layer for 8-class classification
• Input size: 224x224 pixels
• Data augmentation: Random cropping, horizontal flipping

Training Configuration

• Optimizer: SGD with momentum (0.9)
• Learning rate: 0.001 with step decay
• Batch size: 32
• Number of epochs: 24
• Loss function: Cross Entropy Loss

Performance

• The model achieves 100% validation accuracy
• Training completed in approximately 63 minutes
• Uses CUDA for GPU acceleration when available

Explainability

The project includes Grad-CAM visualizations to provide interpretability of model decisions, showing which parts of the dance images the model focuses on for classification.

Requirements

• Python 3.x
• PyTorch
• torchvision
• numpy
• matplotlib
• grad-cam

Usage

1. Training the model:

python CNN.ipynb

2. Generating visualizations:

python resnet18-gradcam-visualization-flat.py

Model Artifacts

• The trained model weights are saved in dance_model.pth
• Grad-CAM visualizations are saved as gradcam_*.jpg

Dataset Organization

• Training data is organized in class-specific folders
• Images are preprocessed and normalized
• Data is split into train, validation, and test sets

Citation

If you use this project, please cite our work: [Project citation to be added]