rusenbb / AIzheimer

We explore the ways to "Unlearn" (you read it right), certain concepts/things while retaining general ability of a generative text-to-image model (stable diffusion 2.1). This project is conducted as a term project for Learning from data class at ITU (Istanbul Technical University)

AIZHEIMER: How was I Drawing?

Overview

AIZHEIMER: How was I Drawing? is a project aimed at exploring the concept of "unlearning" in AI models, specifically focusing on the Stable Diffusion model. The project demonstrates how to selectively forget certain concepts from a pre-trained model using custom attention mechanisms and loss functions.

The offficial introduction video (it is a parody video 😂):

AIZHEIMER: How was I Drawing? ⏯️

Table of Contents

• Overview
• Installation
• Usage
• Project Structure
• Contributors
• License

Installation

To set up the project, follow these steps:

1. Clone the repository:

   git clone [e-mail]:rusenbb/AIzheimer.git
   cd aizheimer

2. Install the required packages: You can run the first cell in the Jupyter notebook to install the necessary packages or use the following command:

   pip install diffusers transformers accelerate torchmetrics[image] torch-fidelity

3. Download the necessary models: Ensure you have the required models from Hugging Face or other sources as specified in the notebook. The models should be automatically downloaded if not locally available. So it should not be a problem.

Usage

1. Run the Jupyter Notebook: Open the aizheimer_main.ipynb notebook in Jupyter and follow the instructions to execute the cells. This notebook contains the entire workflow for setting up the environment, preparing the data, training the model, and testing the results.

2. Configuration: You can configure various parameters such as model_save_path, concept_to_forget, num_image_generate, num_epochs, batch_size, learning_rate, and weight_decay in the notebook.

3. Training: The training process involves using custom attention mechanisms and loss functions to selectively forget the specified concept from the model.

4. Inference: After training, you can test the model by generating images and observing the results to ensure the specified concept has been forgotten.

Project Structure

AIZHEIMER/
├── README.md
├── aizheimer_main.ipynb
└── generated_images/

• README.md: This file.
• aizheimer_main.ipynb: Jupyter notebook containing the main code for the project.
• generated_images/: Directory for storing generated images or training images.

IMPORTANT:

If you decided to not use model generated the images,
you should have training images in the generated_images folder.

Contributors

• Muhammed Ruşen Birben
• Göktürk Batın Dervişoğlu
• Abdulkadir Külçe

We would like to thank our advisor Behçet Uğur Töreyin for his guidance and support throughout the project.

License

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