PyTorch 1.4 Released with improved mobile support

January 15, PyTorch development team released the latest stable version PyTorch 1.4.

PyTorch is a framework that provides algorithm for Tensor computation and neural network that are used in artificial intelligence and reinforcement learning. With GPU acceleration, processing is executed with great speed. Developed by Facebook, it was released at the end of 2018 as open-source software.

PyTorch 1.4 is the latest version, following v1.3 released in October 2019. PyTorch Mobile, which was made open source in v1.3, has been improved, and now you can customize build script at a detailed level. By only including the operators used by the models, library can be optimized, and that will greatly reduce the footprint of the mobile application. When compared with the prebuilt library, customized MobileNetV2 was 40-50% less.

It experimentally introduces new distributed model parallel training. Using distributed RPC framework, it can run functions remotely and refer remote objects without copying the real data.

In addition to Python and C++, it includes experimental support for Java bindings. Based on the interface developed for Android in PyTorch Mobile, it lets you invoke TorchScript models from any Java program. As of now, Java bindings are only available for Linux, and for inference only.

PyTorch domain libraries such as torchvision, torchtext, and torchaudio have been update. For example, torchvision, whose latest version is 0.5, now supports operating environments like quantization, TorchScript, and ONNX.

PyTorch 1.4 is available on the project’s website.