PyTorch, the popular open source machine learning framework, has been updated to version 1.2.0. The new release includes over 1900 fixes that cover JIT, ONNX, distributed learning modes, and performance improvements.
Some changes:
- New TorchScript API Allows it's easy to convert nn.Module (including submodules and methods called in forward()) to ScriptModule.
- Together with Microsoft added full support for ONNX Opset versions 7(v1.2), 8(v1.3), 9(v1.4) and 10 (v1.5). In addition, users can now register their own symbols for exporting custom operations and specify dynamic input sizes during export.
- tensorboard support is no longer experimental.
- Added nn.Transformer module based on article Attention Is All You Need.
- Numerous improvements to the C++ API.
Source: linux.org.ru