Example of RNN (incl. GRU/LSTM) with APIs defined in c_predict_api.h

[Isa-rentacs]

Description

With MXNet 1.3.0, RNN layers can be hybridized. I would like to use those hybridized RNN models from C-api (c_predict_api.h) but couldn't find an example.

More details

I was looking at an issue (How to use LSTM or GRU in C++ using c_predict_api.h?) and in the answer:

@xiaogehh currently the Gluon LSTM and GRU RNN layers cannot be hybridized but the work towards should be soon completed. You can track the effort here: #10873

I believe it's released as a part of MXNet 1.3.0, and now I would like to know how to use them from C++.
If we can have something like the one for CNN it would be very helpful.

[marcoabreu]

FYI: All C-APIs are considered internal APIs that don't follow semantic versioning. They are not intended for external use since they might change at any point in time. We got the Cpp-package which is enduser-facing, so this might rather be a feature request for the C++ package.

[Isa-rentacs]

Thank you @marcoabreu for your answer. Does it mean at this moment loading hybridized RNN models from C++ API is not supported?

[vrakesh]

@mxnet-label-bot [Feature request]

[marcoabreu]

Sorry, I'm not familiar with the c API.

@lebeg @leleamol might be able to answer that question

[lebeg]

@Isa-rentacs In my understanding the implemented RNN layer is implemented in python, so there would be no option to access it from the C API. Maybe this could be a feature request.

[Isa-rentacs]

Thanks @lebeg, I understood this is going to be a feature request. Before going forward, I would like to understand a little bit more.

In the 1.3.0 release note, it says:

Now many dynamic networks that are based on Gluon RNN layers can now be completely hybridized, exported, and used in the inference APIs in other language bindings such as R, Scala, etc.

Do you mean C/C++ is not included in the languages that can import the exported model?

[lebeg]

@Isa-rentacs I'm afraid I don't know the answer. Maybe @szha can help on this?

[leleamol]

@Isa-rentacs The C++ API in cpp-package support methods to load the trained models and parameters.
I have written an example to load a CNN model as mentioned in this PR #12930

Based on issue description, I am inferring that you are planning to use C++ API for inference. Will you be able to share your use case?

[leleamol]

@Isa-rentacs I have submitted the PR #13680 that demonstrates the usage of C++ API in running the inference on RNN models.

The C API are the internal APIs and are not meant for external consumption.

[leleamol]

@mxnet-label-bot add [Pending Requester Info]

[Isa-rentacs]

Hi @leleamol, thank you for working on this, I've added several comments to your PR.

[leleamol]

The example is merged and available in the master. This issue can be closed.

[lanking520]

@Isa-rentacs As explained the PR merged, let's close this issue for now. Please feel free to reopen this thread if you observed more problems with that.