The okosystem and exchange format open neural network exchange (onnx) is in version 1.8 appeared. The release has carried out the serialization of inputs and outputs of the sequence and map data types to ensure operator unit test.
With the serial conveying of the size of elements in make_tensor, users are to ensure that coarse ratios and elements are consistent in a tensor. A highlight for windows users was allowed to be that the current version of onnx contains a new conda package.
Updated training module
Further changes relate to the output loops and the shape inference, which has received some corrections on the node and graphene level. Recently, the module should also be used for models via 2 gb. For the benefit of this change, users can modify the onnx api. The onnx team has worked the modules for training and for the shape inference. For the definition of the gradient operator, the training module has received differentiable tags and a tool that supports developer when storing training information by means of protobuf messages. The graphcall remotes, instead the onnx team has updated ir and graphics of the training module.
Source exchangeable format for ml models
Originally, facebook and microsoft onnx 2017 had to totle to change ml-prerker between machine learning frameworks. The goal is a new standard that allows developers to use their models outside the context in which they created them. Since 2019, the linux foundation has dwelled around the onnx project and continues to develop to establish a manufacturer-independent standard exchange format for machine-learning models. The transaction version onnx 1.7 was published in may 2020 and marked the first step to exchanging ml models during the training.
Details about the current version can be found in github release notes. Further information about the project and its operators will find interested parties on the site of onnx.