PyTorch Translate is a deprecated library for machine translation built on PyTorch, offering sequence-to-sequence model training and export capabilities to Caffe2 via ONNX for production deployment. It targets researchers and engineers working with PyTorch-based NLP models, providing a framework for training and evaluating translation systems.

How It Works

The library leverages the fairseq framework for its core sequence-to-sequence modeling. It allows for training and evaluation of translation models, with a key feature being the export of trained models to Caffe2 graphs using ONNX. This enables deployment in C++ environments, with components like the encoder and decoder exportable separately, and beam search implemented in C++.

Quick Start & Requirements

• Install: python setup.py install after installing PyTorch and fairseq.
• Prerequisites: Linux, CUDA-compatible GPU, GNU C++ compiler (>= 4.9.2), CUDA (8.0 or 9.0 recommended).
• Docker: sudo docker pull pytorch/translate followed by sudo nvidia-docker run -i -t --rm pytorch/translate /bin/bash.
• Full Install: Detailed instructions for installing from source, including Miniconda setup, PyTorch nightly build, and ONNX installation, are provided.
• Docs: Usage examples for training, evaluation, and ONNX export are available within the repository.

Highlighted Details

• Supports training sequence-to-sequence models for machine translation.
• Enables export of models to Caffe2 graphs via ONNX for C++ deployment.
• Includes example scripts for training and evaluating on the IWSLT 2014 German-English task.
• Offers TensorBoard integration for visualizing training statistics.

Maintenance & Community

The project is explicitly marked as deprecated, recommending users switch to fairseq. Contributions are welcomed, with a CONTRIBUTING.md file available.

Licensing & Compatibility

BSD-licensed, allowing for commercial use and integration with closed-source projects.

Limitations & Caveats

The project is deprecated and no longer actively maintained, with users advised to migrate to fairseq. Research models are explicitly noted as works in progress and unsupported.