Does OpenVINO support time-series regression not related to images?

[Polgar__Peter]

Hi,

I searched through the OpenVINO documentation and it seems that OpenVINO works very well with officially supported topologies and models related to computer vision, but what about others? Is there a way that the model mentioned below can be used with OpenVINO (i.e. to convert by Model Optimizer and to use by Inference Engine)?

The model is in .onnx format and has non-image input (time-series data) and consists of the following layers:

1. sequenceInput

2. lstmLayer

3. fullyConnected

4. fullyConnected

5. regressionLayer

From this: https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_Supported_Frameworks_Layers.html#onnx_supported_operators

I concluded that the ONNX section does not mention the sequenceInput, fullyConnected and regression symbols, this means that a model with the above mentioned layers is not supported in any way? If these layers should be custom layers, is it true that custom layers are not applicable to NCS 2?

Any help would be greatly appreciated.

[Sahira_Intel]

Hi Peter,

I apologize for the delay in my response. If your model has layers that are unsupported by OpenVINO, the Model Optimizer will classify those as custom layers.This document describes the workflow for running inference on models with custom layers. Here is a tutorial to implement custom layers on Linux and Windows. I hope this information is helpful.

Best Regards,

Sahira 

[Polgar__Peter]

Thank you very much!

So, if I understand well, any model with supported layers should be convertible with the Model Optimizer, no matter what the topology is and what the type of the input is (any tensor), and custom layers can also be handled.

[Sahira_Intel]

Hi Peter,

Almost - not all topologies are supported by OpenVINO. But if your model is built on a supported topology, and unsupported layers are handled as custom layers then your model will run on OpenVINO.

I hope that was helpful.

Best,
Sahira 

[Polgar__Peter]

Hi Sahira,

Thanks a lot. One more question: As in the case of unsupported layers where there is the custom layer workaround, is there a similar way to handle unsupported (custom) topologies?

Best,

Peter

[Sahira_Intel · Accepted solution]

Hi Peter,

There is not a workaround to handle unsupported network topologies. You can visit the OpenVINO documentation here for a comprehensive list of supported topologies for each Framework. 

Best Regards,

Sahira