Skip to content

Style Transfer with Neural Network

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

Style Transfer with Neural Network

README.md

🎨 Style Transfer with Neural Networks 🖼️

Welcome to the Style Transfer with Neural Networks project! In this project, we explore the application of neural networks to perform artistic style transfer, allowing us to blend the artistic style of one image with the content of another.

Table of Content

  1. Project Overview
  2. Setup & Installation
  3. File Structure
  4. How to Run
  5. Sample Run
  6. Concepts Behind the Project
  7. Examples of Generated Images
  8. Technologies Used
  9. Parameters & Tuning
  10. License
  11. Contact

📚 Project Overview

Style transfer is a technique in deep learning that involves using Convolutional Neural Networks (CNNs) to separate and combine the style of one image with the content of another image. This project implements a neural style transfer model using PyTorch and pre-trained VGG19.

Key Features:

  • Use of pre-trained neural networks for feature extraction (VGG19)
  • Content loss and style loss based on Gram matrices
  • Supports custom content and style images
  • Optimization with L-BFGS optimizer for faster convergence

🛠️ Setup & Installation

To get started with this project, you'll need to install the following dependencies:

pip install torch torchvision matplotlib Pillow

File Structure:

File/Folder Description
Style Transfer with Neural Networks.ipynb The main notebook implementing style transfer
images/ Directory containing sample content and style images
output/ Folder to store the generated images
requirements.txt List of dependencies

🚀 How to Run

  1. Clone the repository and navigate to the project folder:

    git clone https://github.com/yourusername/style-transfer-project.git
cd style-transfer-project

  2. Run the Jupyter Notebook to perform style transfer on your custom images. You can upload your content and style images directly in the notebook.

  3. Adjust parameters like content_weight, style_weight, and num_steps to tweak the results according to your preferences.

  4. View and save the generated images in the output/ folder.

📸 Sample Run

  • Content Image: Defines the structure and objects in the final image.
  • Style Image: Provides the textures, colors, and patterns.
Content Image Style Image
Content Image Style Image

🔬 Concepts Behind the Project

  • Content Representation: Extracted from deeper layers of the network to capture high-level structures in the image.

  • Style Representation: Captured using the Gram matrix of feature maps, representing correlations between different feature maps.

  • Optimization: The neural network optimizes a noise image to minimize both content loss and style loss, blending the content and style.

🧠 Technologies Used

  • Python 🐍
  • PyTorch for deep learning
  • Jupyter Notebook for interactive coding
  • Matplotlib for visualizations

📊 Parameters & Tuning

You can adjust the following parameters to control the output of the style transfer model:

Parameter Default Value Description
content_weight 1e5 Weight for the content loss
style_weight 1e10 Weight for the style loss
num_steps 300 Number of optimization steps
learning_rate 0.01 Learning rate for the optimizer

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

💬 Contact

If you have any questions, feel free to reach out to me at [alolikabhowmik72@gmail.com]