GPT-2 - Training and Inference

This repository is building a GPT-2 model from scratch, coding the architecture and providing separate modules for both inference and training.

Overview

GPT-2 implementation its architecture, training it on external data, and visualizing its attention mechanisms on layer and head. The repository includes separate modules for inference and training of gpt2.

Features

Building GPT-2 Architecture

• From Scratch Implementation: The GPT-2 architecture, including its essential components such as attention mechanisms, layer normalization, and positional embeddings, is built from scratch.
• Drawing Inspiration: The implementation closely parallels the Hugging Face Transformer library to ensure compatibility and best practices.

Inference Module

• Custom Inference Module: A fully functional inference module has been developed from the ground up.
• Parameter Loading: 124 million parameters are loaded from the Hugging Face library to ensure compatibility with standard GPT-2 configurations.
• Comparison Analysis: The custom-built GPT-2 model’s inference results are compared against the Hugging Face GPT-2 model.

Key Insights:

1. Result Parity: The outputs from the custom GPT-2 model match those of the Hugging Face GPT-2 model.
2. Pipeline Discrepancy: While the underlying models produce similar results, text generated using the Hugging Face pipeline differs due to additional preprocessing and sampling strategies.

Training Module

• External Data Integration: External datasets are loaded and preprocessed to train the custom GPT-2 architecture effectively.
• Custom Training Loop: A training loop is implemented to fine-tune the model and evaluate performance iteratively.

Attention Visualization Module

• Understanding Attention: A dedicated module visualizes attention mechanisms for better interpretability of the model.
• Token Mapping: This module highlights which tokens in the input sequence attend to others, providing insights into how the model processes information.

Model Architecture

The GPT-2 model is designed using the following parameters:

• vocab_size: 50257
  • Size of the vocabulary, representing all possible tokens.
• n_embed: 768
  • Dimensionality of token embeddings and hidden states.
• block_size: 1024
  • Maximum context length for processing sequences.
• n_blocks: 12
  • Number of transformer blocks in the architecture.
• n_heads: 12
  • Number of attention heads in each block.

How to Use This Repository

1. Inference:

   • Load the pre-trained parameters using the provided scripts.
   • Use the inference module to generate text based on input prompts.

2. Training:

   • Prepare external datasets for fine-tuning the architecture.
   • Run the training scripts to update model weights.

3. Attention Visualization:

   • Use the visualization module to analyze token attention for input examples.