Tap steps to explore sequence:
Select a step circle above to explore the Transformer flow.
PyTorch Custom Architecture
Sequence-to-Sequence Transformer
This project was an educational deep-dive into the "Attention Is All You Need" architecture. Instead of using pre-built library models, this implements the core encoder-decoder transformer blocks, multi-head attention, and masking entirely from the ground up in PyTorch.
- Built completely from scratch using standard PyTorch nn.Module
- Implements full Multi-Head Attention mechanisms
- Positional Encoding layer for sequence awareness
- Trained for language translation tasks
Computer Vision
Vision Transformer (ViT)
Adapting transformer architecture for images. This project breaks down images into patches, linearly embeds them, and feeds them into a standard transformer encoder, demonstrating how sequence models can perform state-of-the-art computer vision tasks without standard Convolutional Neural Networks (CNNs).
- Custom patch embedding layers
- Self-attention for image classification
- Improved baseline accuracy from 10% to 80%+
- Built for academic research in IS Masters program
"Love your enemies and do good to those who hate you."
T5
"Love ye your enemies, and do good unto them which hate you."
"He went to the store and bought bread."
T5
"And he went unto the market, and purchased bread thereof."
"Do not be afraid, for I am with you always."
T5
"Fear not; for I am with thee alway, even unto the end."
Hugging Face Fine-Tuning
T5 Text Styling Model
Using transfer learning, I took a pre-trained T5 model and fine-tuned it on a parallel corpus of modern English and King James Version text. The resulting model can take standard sentences and re-style them with biblical vocabulary and sentence structure.
- Fine-tuned Google T5 (Text-to-Text Transfer Transformer)
- Translates modern English into KJV-style prose
- Utilizes Hugging Face transformers pipeline
- Deployed inside Jupyter Notebooks for testing
Tap steps to explore sequence:
Select a step circle above to explore the U-Net architecture.
Latent Space Generation
Diffusion Generative Models
An exploration into generative AI. This project covers the mathematics and implementation of diffusion models, starting from adding Gaussian noise to an image and training a U-Net to predict and subtract that noise, eventually creating novel images from pure noise.
- Built a from-scratch U-Net architecture
- Experiments with Hugging Face UNet2DModel
- VAE-based latent space processing
- Forward noise addition and reverse denoising loops