Facebook announced the availability of PyTorch 1.1 as part of its developer conference F8. The new release of the open source deep learning library, which has been open since 2017, is expected to deliver better performance and new tools to better understand and visualize the machine learning models.
In addition, the developers have provided new APIs on Facebook. The predecessor release, PyTorch 1.0, was released in December 2018 and is now being used by companies such as Microsoft, Toyota and Airbnb, as the social network outlined in the 1.1 announcement.
JIT compiler now more mature
The highlight of the new release is probably the once again revised JIT compiler. In version 1.0, an alpha version of the compiler had landed, which probably did not work much faster than the basic mode of PyTorch. This has now been improved with the new edition of the JIT compiler, especially since it is likely that many more programming language concepts from Python to PyTorch can be implemented. The JIT compiler is able to determine at runtime how to generate the most efficient code.
PyTorch 1.1 also provides the ability to split the neural networks into GPUs. Previously, PyTorch allowed developers to split training data into processors for data-parallelism. The sharding of the models now also allows an instruction parallelism with which the networks can now also achieve “Multiple Instruction, Multiple Data” (MIMD) techniques.
Read More Stories: Social Network Analysis: In 50 years more dead than lively users on Facebook
With the command “from torch.utils.tensorboard import SummaryWriter”, the library TensorBoard now supports natively, a web application for the review and understanding of training routines and graphics. Among the new APIs are support for Boolean tensors, while those for custom recurrent neural networks have been revised.
Finally, it has worked well with the community to promote projects and tools that help ML professionals with their needs, ranging from improved model understanding to auto-tuning with AutoML methods. With the release of Ax and BoTorch Facebook has therefore opened core algorithms. BoTorch is a framework based on PyTorch to enable Bayesian optimization for the sequential optimization of black box features. In turn, Ax is a machine-learning adaptive experiment management platform that systematically explores large configuration spaces to customize models, infrastructure, and machine-learning products.
Read More Stories: OpenBSD 6.5 with new stack protector
Further information on the new PyTorch release and further developments from the PyTorch ecosystem can be found in the blog announcement. If you want to get deeper into PyTorch, you will get more detailed information on the project website.
Read More Stories: Apple At WWDC 2019 Preview: New iPads, MacBooks and iOS 13