DarkVision

DarkVision is a deep learning project focused on image classification using PyTorch and ResNet18 architecture. The project implements various loss functions and provides tools for training and testing image classification models.

Project Structure

DarkVision/
├── dataset/                 # Directory containing training and validation datasets
├── model/                   # Directory for model-related files
├── final_results_NV/        # Directory for storing final results
├── checkpoint.pth           # Saved model checkpoint
├── training_output.txt      # Training process output
├── utils.py                 # Utility functions and data loading
├── test.py                  # Testing script
├── test_base.py             # Base testing functionality
├── train.py                 # Training script
├── loss.py                  # Custom loss function implementations
├── model.py                 # Model architecture definition
├── data.ipynb               # Jupyter notebook for data analysis
└── LICENSE                  # Project license file

Features

• ResNet18-based image classification
• Multiple loss function options:
  • Focal Loss
  • Mean Squared Error (MSE)
  • Cross Entropy Loss
• Custom dataset handling
• Training and validation pipeline
• Model checkpointing
• GPU support (if available)

Prerequisites

• Python 3.x
• PyTorch
• torchvision
• pandas
• numpy
• PIL (Python Imaging Library)

Installation

1. Clone the repository:

git clone https://github.com/michaelborek/DarkVision.git
cd DarkVision

2. Install the required dependencies:

pip install torch torchvision pandas numpy pillow

Dataset Preparation

The project expects datasets in CSV format with the following structure:

• filepath: Path to the image file
• label: Class label for the image

Place your training and validation datasets in the dataset/ directory:

• train_dataset.csv
• val_dataset.csv

Usage

Training

To train the model, use the following command:

python train.py --loss [loss_type]

Available loss types:

• focal: Focal Loss
• mse: Mean Squared Error
• cross_entropy or ce: Cross Entropy Loss

Example:

python train.py --loss focal

Testing

To test the model, use:

python test.py

Model Architecture

The project uses a ResNet18 architecture with the following modifications:

• Pretrained weights from ImageNet
• Customizable number of output classes
• Modified final fully connected layer

Training Parameters

• Number of epochs: 8
• Batch size: 128
• Learning rate: 0.001
• Optimizer: Adam

Results

Training progress and results are saved in:

• training_output.txt: Training metrics and progress
• checkpoint.pth: Model weights and training state
• final_results_NV/: Directory containing final evaluation results

License

This project is licensed under the terms specified in the LICENSE file.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Contact

For any questions or suggestions, please open an issue in the repository.