Hi! I’m one the maintainers of tiny-dnn and we’d like to know the position of this community regarding import/export models from/to other frameworks?
We are also in touch with the Khronos group to help with their initiative to define a proper Neural Network Exchange Format (NNEF) which could make things easier. More details
Meanwhile, do you have already any internal data structure to read the models definition?
We’re open to discussion about formats for sharing parameters and model definitions between libraries. If there’s going to be a larger initiative with some of the frameworks adopting a new format, we’ll likely join and try to help.
Right now our serialization is very torch specific and we didn’t design it having in mind interoperability with other software. If you need to export model weights I encourage using state_dict and some more common formats like HDF5 or protobuf.
is there a suggested new format to export in pytorch for use in other framework and custom compilers?