DeepJetCore 2.0
First 2.0 release.
The following things changed w.r.t. 1.X releases:
-
The underlying data format changed drastically and is incompatible with 1.X (conversion scripts provided)
-
It is built entirely in compiled C++ modules now
-
It fully supports ragged tensors, ragged in the first dimension, which should be sufficient for most use cases.
-
Full ragged support will have to wait for keras patches that should come soon. But data format compatibility will be guaranteed through the 2.X releases.
-
The code has been cleaned up drastically, but will be further cleaned up in next releases.
-
The user interface has been simplified, and changed. In particular, the TrainData_XX classes are more transparent; for more information, please refer to the example generated by createSubPackage.py
-
A slim C++ interface/library for the data structures can be extracted (extractDjcCppInterface.py), such that the data can be written from any other C++ framework without the need of the full DeepJetCore package and its dependencies.
-
... and many more, for more details check the diff