Enhanced Swin Transformer Model

A deep learning project implementing an enhanced Swin Transformer model trained on the Tiny-ImageNet dataset.

Project Structure

.
├── debug_visualizations/         # Results after executing the code
├── tiny-imagenet-200/            # Complete dataset directory
├── train/                        # Subset of the dataset for training
├── val/                          # Subset of the dataset for validation
├── enhanced_swin_model.py        # Swin Transformer model implementation
├── requirements.txt              # Project dependencies
├── tiny-imagenet-downloader.py   # Script to download the dataset
├── train_small.sh                # Script for training on a small subset
├── traning-run-script.sh         # Script for full dataset training
└── training-script.py            # Main training code

Model Architecture

The Enhanced Swin Transformer implemented in enhanced_swin_model.py builds upon the original Swin Transformer architecture with several improvements:

• Hierarchical feature representation
• Shifted window-based self-attention
• Efficient computation with linear complexity to image size
• Improved patch merging and embedding layers

Future Enhancements

The project roadmap includes:

1. Data Augmentation: Implementing diverse data augmentation techniques to improve model generalization capabilities.

2. Robustness Enhancements: Adding resistance against common image modifications such as JPEG compression and noise to increase real-world applicability.

3. Progressive Training Strategy: Implementing curriculum learning to train the model progressively from simple to complex examples.

4. Adversarial Training: Incorporating adversarial training methods against steganalysis detectors to improve model robustness.

Results

Training results and visualization outputs can be found in the debug_visualizations directory after running the training scripts.