Paper-to-Code Implementations

Overview

This repository is a collection of implementations of influential machine learning and deep learning papers. Each model is translated from theory to code, offering practical, executable versions of pioneering architectures. This project serves as both a learning exercise and a portfolio piece to demonstrate understanding and application of advanced models across various tasks, from generative modeling to object detection and classification.

Repository Structure

Each model has its own directory containing the following:

• model.py: Model architecture.
• train.py: Training loop and setup.
• inference.py: Inference script for generating predictions.
• utils.py: Utility functions specific to each model.
• README.md: Detailed instructions and context for the implementation, including setup, usage, and results.

Models Included

Model	Task	Dataset	Paper Reference	Highlights
CycleGAN	Image-to-Image Translation	Cezanne2Photo; Car2CarDamage	CycleGAN	Unpaired image translation
DCGAN	Image Generation	MNIST	DCGAN	Deep Convolutional GAN for realistic image generation
ESRGAN	Image Super-Resolution	Pre-trained	ESRGAN	Enhanced super-resolution GAN with high-quality outputs
PointNet	3D Object Recognition	ModelNet40	PointNet	Processing 3D point clouds
ProGAN	Image Generation	CelebA	ProGAN	Progressive growing GAN for high-resolution images
VAE	Image Generation	MNIST	VAE	Variational Autoencoder for learning complex distributions
VQGAN	Image Generation	Oxford Flowers	VQGAN	Combines GANs with vector quantization for high-quality images
ViT	Image Classification	FoodVision Mini	ViT	Vision Transformers for image recognition
WGAN / WGAN-GP	Stable GAN Training	MNIST	WGAN	Improved GAN training stability
YOLOv3	Object Detection	Pascal VOC	YOLOv3	Real-time object detection model
pix2pix	Image-to-Image Translation	Sat2Map	pix2pix	Paired image translation
vgg_lpips	Image Similarity	Custom	LPIPS	Learned Perceptual Image Patch Similarity for comparing images

Setup Instructions

Each model requires a specific environment configuration. Please refer to the requirements.txt or environment.yml file within each model’s directory for dependency information.

To set up a conda environment with the dependencies, use:

conda env create -f environment.yml
conda activate paper-to-code

Alternatively, install dependencies via pip:

pip install -r requirements.txt

How to Use

Each model directory includes instructions for training and inference:

1. Training: To train a model, navigate to the corresponding directory and run:

   python train.py

2. Inference: For inference on new data, use:

   python inference.py --input_path <path_to_input> --output_path <path_to_output>

3. Visualization: Each model includes visualization tools to view outputs, such as generated images or bounding boxes.

Future Directions and Potential Enhancements

1. Documentation & Notebooks: Adding interactive Jupyter notebooks to walk through each model’s training and inference process.
2. Additional Models: Expanding the repository to include:
   • UNet for medical image segmentation.
   • ResNet or EfficientNet as baseline image classifiers.
   • DETR for object detection.
   • StyleGAN for more sophisticated image generation.
3. Comparison Table: Summarizing performance metrics, training times, and notable findings in a central location to showcase the breadth of experimentation and model capabilities.

Contributing

This repository is open for contributions. If you have suggestions or improvements, please submit a pull request or reach out with feedback.

Contact

For any questions or feedback, please feel free to contact me via Email or at LinkedIn.