The open-source machine learning framework PyTorch is tackling production usage in its latest release. PyTorch 1.2 features an update to the TorchScript environment. TorchScript enables users to create serializable models from PyTorch code and can be saved from a Python process.
The new improvements are designed to make it easier to ship production models, expand support for ONNX formatted models and enhance support for Transformers. Improvements include support for the subset of Python in PyTorch models, and a new API for compiling models to TorchScript.
The release adds full support to export ONNX Opset version 7, 8, 9, and 10.
The team also announced TensorBoard has graduated from its experimental phase.
Other features of the release include domain API library updates. Torchaudio 0.3, its machine learning library for signal processing functionality, has been released with Kaldi compatibility, a new tutorial, a focus on standardization and two new functionals. Torchtext 0.4, its natural language processing library, comes with popular supervised learning baselines with “one-command” data loading. The datasets include AG_NEWS, SogouNews, DBpedia, YelpReviewFull and YahooAnswers. Torchvision 0.4, its video library, includes standard video datasets, IO primitives for reading and writing video files, support for arbitrary encodings and formats, and reference training scripts.
