The ecosystem and exchange format Open Neural Network Exchange (ONNX) has been released in version 1.8. The release has revised the serialization of inputs and outputs of the sequence and card data types to enable operator unit testing.

With the serial test the size of elements in make_tensor should allow users to ensure that proportions and elements in a tensor are consistent. A highlight for Windows users should be that the current version of ONNX contains a new conda package.

Updated training module Further changes affect the output loops and the shape inference, which has received some corrections at the node and graph level. Recently, the module should also be able to be used in models over 2 GB. To take advantage of this change, users can modify the ONNX API. The ONNX team has revised the modules for training and shape inference. For the definition of the gradient operator, the training module has been given differentiable tags and a tool that supports developers in storing training information using Protobuf messages. The GraphCall is omitted, instead the ONNX team has updated the IR and graphics of the training module.

Open source exchange format for ML models Facebook and Microsoft originally initiated ONNX 2017 so that ML practitioners can switch between machine learning frameworks. The goal is a new standard that allows developers to use their models outside of the context in which they were created. Since 2019 the Linux Foundation has been taking care of the ONNX project and developing it further in order to establish a manufacturer-independent standard exchange format for machine learning models. The previous version ONNX 1.7 was released in May 2020 and marked the first step towards exchanging ML models during training.

Details for the current version can be found in the release notes on GitHub. Further information on the project and its operators can be found on the ONNX website.

(sih)