zfit / zfit-development

The developement repository for zfit with roadmaps, internal docs etc to clean up the issues
0 stars 2 forks source link

Copy and Dumping/Serialize #6

Open jonas-eschle opened 5 years ago

jonas-eschle commented 5 years ago

Loading and dumping PDFs and objects to YAML (or similar)

This allows to

this is somewhat orthogonal to using a ML native format, which could be better for preservation (such as for a paper), since it usually cannot be modified anymore. ONNX and the TF native graph dumping are good candidates.

For the data: the ASDF format seems promising. It combines hr specs (YAML) with serialized data.

In order to print nice representations and to dump/load any object, it has to know about it's own representation.

This is also related to a correct copy of the object: once an object know how to dump and load itself, a copy is basically for free

The current status is an experimental implementation. Each object has to be able to create itself, it can use

which dumps itself into an intermediate representation (a whole new set of classes). From this, either a copy or a serialization can be made. This additional layer allows to decouple the serialization from the actual object