PyTorch's forgotten translation library, now a museum piece

What it does PyTorch Translate was a machine translation toolkit built on top of fairseq. It trained sequence-to-sequence models and, more specifically, exported encoder/decoder components to Caffe2 graphs through ONNX for C++ production inference. Beam search stayed in C++ land, with promises of full export that never arrived.

The interesting bit The project existed mainly to solve a deployment problem: how to get PyTorch research models into Facebook’s Caffe2 production infrastructure. The README’s “near future” promises about exporting beam search and adding more model support now read as historical fiction — the entire repo was deprecated in favor of fairseq itself.

Key highlights

• Built as a thin layer over fairseq; models interoperated between both codebases
• Exported encoder and decoder separately to .pb Caffe2 graphs via ONNX
• Provided Docker images with CUDA 8/9 and elaborate conda-based install scripts
• Included IWSLT 2014 de-en pretrained model and ensemble training setup
• TensorBoard logging supported through a separate train_with_tensorboard.py script

Caveats

• Explicitly deprecated; README opens with a migration notice pointing to fairseq
• Linux and CUDA-only; the install process involved manual NCCL2 downloads and library path wrangling
• Research models in pytorch_translate/research were explicitly unsupported works in progress

Verdict Worth studying if you’re maintaining legacy Caffe2/ONNX export pipelines or writing software archaeology papers. Everyone else should use fairseq directly, as the maintainers themselves recommend.