GNA Plugin: Inferring CNN1D with device GNA2.0 does not work

[tea6329714]

[Zulkifli_Intel]

Hello tea6329714.

Thank you for reaching out to us.

I ran a test on both of your python scripts, here are the results from my test.

For NN1_CNN1D_GNA2.0Infer_NotWork.py

For NN2_CNN1D_GNA2.0Infer_NotWork.py

As you can see I received errors for both tests.

Please share with us your CPU specifications and the models' precision.

Sincerely,

Zulkifli  

[tea6329714]

Hi Zulkifli,

Thanks for your reply.

The reason of error for your both test is that you may use hardware GNA3.0 to infer NN1 & NN2.

NN1 & NN2's filter number (36684 & 9216)  are not up to the  GNA3.0 specifiaction (filter number <8192).

#{CPU & GNA information}

#{IR information: } precision is float 32

#{GNA3.0 spec}

[tea6329714]

Hi Zulkifli,

I can duplicate the error when I use another computer with CPU (12th Gen Intel(R) Core(TM)) i7-12700H).

I use the same virtual environment (openvino 2021.4) to run the script and get the same error message .

Finally, i use virtual environment (openvino 2022.1) to regenerate IR & infer scripts ,

and the script (NN2_CNN1D_GNA2.0Infer_NotWork.py) works.

So I update the attaced package (IR was complied from openvino 2022.1).

[Zulkifli_Intel]

Hello tea6329714.

I apologize for the late reply.

I tested your models with the Benchmark Python Tool and got the same result as yours.

We are still investigating this and will get back to you with the result of our investigation.

Sincerely,

Zulkifli 

[Zulkifli_Intel]

Hello tea6329714.

We are still investigating this matter and our first assumption is that the Tracing_NN1_CNN1D model may contain a layer that is not unsupported by the GNA plugin. Here is the list of Supported Layers by GNA Plugin that you can refer to.

Sincerely,

Zulkifli 

[tea6329714]

Hi Zulkifli,

Thanks for your reply.

We do tests to get CNN1D’s 3 limitations, but we can’t confirm the correctness.

We need intel teams to help confirm these 3 limitations (limitation 1: cnn1d's input element < 65528 for GNA.  Limitation 2: cnn1d's input channel < 256 for GNA. Limitation 3: cnn1d's output channels <8192 for GNA3.0).

About fail running of Tracing_NN1_CNN1D, We guessed that input channel of  Tracing_NN1_CNN1D model exceeded the limiation of CNN1D input channel (384<256). 

Tip: our tests and result can be got from attached files.

tea6329714

[Zulkifli_Intel]

 Hello tea6329714,

Thank you for your patience,

We verified all of your tests and we would like to inform you that the GNA 3.0 limitations that you found are correct.

Thank you for asking this question, we'll update the documentation once we received the approval from the developer side.

Sincerely,

Zulkifli

[Zulkifli_Intel]

Hello tea6329714,

Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.

Sincerely,

Zulkifli