TorchLingo

TorchLingo is an educational PyTorch library for Neural Machine Translation (NMT). Designed for students and instructors, it provides a clean, well-documented implementation of the Transformer architecture for learning and experimentation.

Features

• 🎓 Educational Focus: Clean, readable code designed for learning
• 🔄 Transformer Architecture: Full encoder-decoder implementation with multi-head attention
• 📝 SentencePiece Tokenization: BPE and Unigram subword models
• 🔁 Back-Translation: Data augmentation for improved translation quality
• 🌍 Multilingual Support: Train a single model for multiple language pairs
• 📊 TensorBoard Integration: Monitor training progress in real-time

Installation

pip install torchlingo

For development:

pip install torchlingo[dev]

Working from a clone

The example corpus and the pretrained checkpoint are stored in Git LFS. Install it before cloning, or the two files arrive as short text pointers instead of data and the tutorials that read them will not run:

git lfs install
git clone https://github.com/BYU-Matrix-Lab/torchlingo.git

Already cloned without it? git lfs install && git lfs pull fetches them.

Documentation

For full documentation, tutorials, and API reference, visit:

• Getting Started Guide
• API Reference

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the GNU Affero General Public License v3.0 - see the LICENSE file for details.

Acknowledgements

This project was initially developed by Josh Christensen as part of his undergraduate work at BYU.

"I hope that TorchLingo will be a valuable resource for students learning about neural machine translation, and that they will consider improving this project and the entire world with the knowledge they gain."
— Josh Christensen