Facebook recently launched a new version of its deep learning framework PyTorch. The update goes under the name PyTorch 1.4 and includes limited Java support, adjustments to the mobile version of the framework and various upgrades for the audio, vision and text libraries.
PyTorch is one of the most popular machine learning frameworks currently used by developers and researchers. Initially it was released for desktop and last year it added mobile support for iOS and Android devices with PyTorch 1.3.
Mobile upgrade PyTorch 1.4
As mentioned before, Facebook has further tinkered with the mobile version of the framework with version 1.4. In this way, developers can start using the latest version of PyTorch Mobile with sophisticated library modifications. This makes it possible to better optimise the size of libraries. In early tests, it was possible to generate a version of MobileNetV2 that was 50 percent smaller than the original model.
Torch library modifications
PyTorch 1.4 also comes with a domain library upgrade with torch-scriptable support for all models. This makes it easier to get started in non-Python environments. Additionally, with the new version ONNX, torchvision gets support for all models.
The torch audio library also has new features, including new filters and interactive speech recognition (now experimental). On the other hand, the torchtext library can now connect to the enwik9 unsupervised learning data set.
Limited Java support
The update also includes a framework for distributed model parallel training and limited Java support for PyTorch inference based on the PyTorch Mobile for Android interface. At the launch of PyTorch 1.4 however, the experimental feature will only be available for Linux and inference. At the moment PyTorch supports Python and C++.
Last month Facebook also released some new features, including the image and video classification framework called PyTorch Elastic.