Facebook has released what it’s modestly calling “a large feature update” to PyTorch with a clutch of fresh tools.
Five months after PyTorch 1.0, Facebook announced an iterative release targeting deep coders using the this open-source, natural-language Machine Learning library.
Included in version 1.1 is native support for virtualisation and model debugging in TensorFlow’s virtualisation toolkit TensorBoard.
TensorBoard provides a range of capabilities including tracking and visualisation metrics, model graph and histograms for analysis of training runs and graphs. PyTorch 1.1 will now work with a simple “from torch.utils.tensorboard import SummaryWriter” command.
Changes to Just-in-Time compilation (JIT) include bug fixes and improved support for TorchScript in dictionaries, user classes and attributes.
Support for new APIs has been added, including Boolean tensors and custom Recurrent Neural Networks. Specifically, the newly supported APIs are intended to make it easier for developers to write their own, fast custom RNNs with TorchScript without needing to write specialised CUDA kernels to achieve levels of performance found with supplied RNNs. You can find out more here.
Facebook also promised improved performance for common models including CNNs, and support for multi-device modules, including splitting models across GPUs while still using Distributed Data Parallel.