Issue with Deploying Trained Model to Akida Hardware (AKD1000)

Help
#1

Hi all,

I encountered an issue when deploying a trained model to Akida hardware (AKD1000). Here’s a summary of my workflow:

  1. Model Creation and Training: I created and trained the model using TensorFlow.
  2. Profiling: To profile the model, I used device='brainchip-akd1000'.
  3. Deployment: For deployment, I set the target to Zip, and I also applied quantization for hardware compatibility. I used the below code
    deploy_bytes = ei.model.deploy(model=model_keras,
    model_output_type=model_output_type,
    model_input_type=model_input_type,
    representative_data_for_quantization=X_train,
    deploy_target=‘zip’)
    Unfortunately, this returned an error.
    Could not deploy:

get_pretrained_model_info did not return model details.
Has anyone experienced a similar issue or knows how to fix it? Any guidance would be appreciated!

Thanks in advance for the help.

Rashedul

#2

Hi @Rashedul

Let me ask @mateusz and @jbuckEI if they know, they are best to answer this.

Thanks for the report, hopefully we can get some troubleshooting steps.

Best

Eoin

#3

Hi @Rashedul
Could you clarify your workflow a little bit?
As I understand, you trained your model off the Edge Impulse Studio, then you uploaded the tflite file (or Keras?) to the Studio and ran a profiling against AKD1000?
Are you sure your model is compatible with Akida? This NPU has constraints about the layers, their parameters and other things. Please refer to Akida documentation for details.
My advice is to create and train model using our Studio (easy solution) or Akida Python framework. That way you can be sure that your model is compatible with AKD1000 NPU.

Regards,
Matt