pkraman06/PyTorch_Transformer_model

🌐 Seq2Seq Transformer English-to-Spanish Translator

An interactive web application deployed on Hugging Face Spaces that demonstrates a Sequence-to-Sequence (Seq2Seq) Transformer model built from scratch using PyTorch and served via Streamlit.

The application automatically builds custom word-level vocabularies from a training dataset, handles variable-length sequence padding, trains a classic multi-head attention Transformer network, and performs auto-regressive decoding for real-time translation inference.

🚀 Features

• —Custom Tokenization Pipeline: Processes raw text data, builds independent source and target vocabularies, and converts sentences into tensor indices with <SOS> and <EOS> boundaries.
• —Transformer from Scratch: Implements Sinusoidal Positional Encodings and a standard PyTorch Multi-Head Attention Transformer (nn.Transformer) with proper causal masks (tgt_mask) to prevent lookahead during training.
• —On-the-Fly UI Training: Automatically fits the model to the provided dataset on the initial load and caches the trained weights to provide instant inference for users.
• —Streamlit Web Interface: A clean, user-friendly text field interface allowing anyone to input English sentences and see Spanish translations instantaneously.

📂 Project Structure

To run properly on Hugging Face Spaces, ensure your repository contains the following files in the root directory:

├── app.py               # Main Streamlit web application & PyTorch script
├── requirements.txt     # Python runtime dependencies
├── data.csv             # Custom English/Spanish translation dataset
└── README.md            # Project documentation (this file)