BabyGPT: A Transformer from Scratch

BabyGPT is a Python project that implements a simple Transformer model (~17M parameters) from the ground up using TensorFlow and is based on the "Attention is All You Need" paper.

This project demonstrates how a Transformer can be constructed to learn and generate text based on a given input corpus, mimicking the style of the provided dataset.

Features

• Full implementation of the Transformer architecture from scratch.
• Capable of text generation by learning from a custom text file.
• Easy to train using TensorFlow.
• Well-structured codebase to facilitate understanding of the Transformer architecture.

Model Overview

The Transformer model adheres to the architecture outlined in the paper, customizing it for the text generation, which includes:

• Token Embedding
• Positional Encoding
• Multi-Head Self-Attention Mechanism
• Decoder Stack

The model can be trained on a text input file, learning to generate coherent text based on the patterns and structure observed in the data.

Installation

To set up the project, follow these steps:

1. Clone the repository:

   git clone https://github.com/pietrovitagliano/BabyGPT.git
   

2. Install the dependencies:

    pip install -r requirements.txt
   

3. Train the model: Run the model_trainer.py script to train and save the model. This script is designed to load an existing model if available, allowing for continued training across multiple sessions without losing previous progress.

4. Text Generation: Use the text_generator.py script to load the trained model and use it. Simply provide a starting phrase and the model will generate text in the learned style.

Licence

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