DAAL neural network (de)serialization

[miller__steve]

Hi,

I noticed from discussions in the past that DAAL neural network models did not support (de)serialization using the data_management::InputDataArchive/OutputDataArchive classes. Can anyone comment on whether or not this has now been implemented, as I have tried this and not been able to make it work.

If it's not been done, then should we assume the only way to do this is the technique used in the example neural_net_predict_dense_batch.cpp?

Thanks

[Gennady_F_Intel]

Hi,  yes, that's correct  assumption. The DAAL neural network doesn't support (de)serialization mode and there is only one way to do this is the technique used in the example neural_net_predict_dense_batch.cpp.

[miller__steve]

Thank you for your reply.

Can you comment on whether this issue is part of the current development roadmap? It would be very useful to enable re-use of previously trained models.

Steve.

[Gennady_F_Intel]

Hello Steve, thank you for your questions and we agree that this reasonable feature to go. 

We cannot share some concrete dates when it may be available out of the box - we are still continue analyzing and prioritizing the possible future extensions of daal’s functionality. 

How much today's lack of serialization is critical to your applications?

thanks, Gennady

[miller__steve]

Hi,

Thank you very much for your reply. I can proceed with my use of DAAL, which is a great library to use.

The serialization would be a useful feature in the long term though - I think it's worth comparing with some other libraries where this is a relatively simple operation (e.g. tensorflow). Not quite a single function call, but fairly straightforward. 

Many thanks for your time,

Steve.