Serialization / deserialization of optimizers and computing services

[jph11]

Hi! Would it be possible to add specific functions for the serialization of the optimizers and computing services? At this moment, the serialization of the whole neural network is supported and working fine. However, we'd need something similar for the optimizers and CS in order to fully share them with the worker nodes that make up a distributed training.

The serialization and deserialization of the whole neural network takes place this way (Python syntax):

- eddl.import_net_from_onnx_string(self.model)
- eddl.serialize_net_to_onnx_string(model, False)

We'd like to have something similar for the optimizers. A suggested syntax:

- eddl.import_optimizer_from_onnx_string(optimizer_string)
- eddl.serialize_optimizer_to_onnx_string(eddl.sgd(0.01, 0.9))

And also for the computing services:

- eddl.import_cs_from_onnx_string(computing_service_string)
- eddl.serialize_cs_to_onnx_string(eddl.CS_CPU(mem="full_mem"))

[jonandergomez]

Dear jph11,

Please, could review the following C++ example still in the develop branch? https://github.com/deephealthproject/eddl/blob/develop/examples/onnx/4_utils_serialization.cpp

I guess this is what you are asking but for Python, isn't it?

Could you please confirm? Then we will manage to include such methods in the Python API (PyEDDL) in the next release in the master branch.

Regards,

Jon