Hello, I’m having an issue with the BYOM feature. I am doing a 3-class audio classification project, using Mel-spectrograms as input features to my model. I have a pretrained (.tflite format) CNN model, and it is successfully uploaded in the “Upload model” tab. My goal is to quantize and deploy the model to Arduino Nano 33 BLE Sense. After saving the model, in the deployment tab, I don’t see a button to quantize the model (I want to reduce its size and reduce inference time, but only the EON optimizer is offered). Also, I want to see the performance of the model on its testing set when deployed to the MCU. I uploaded a chunk of the testing data using Python SDK and then classified all of it, but then it was evaluated using the float32 model. I’m wondering if there is a way to quantize it and evaluate on the testing set? Also, when uploading the test set (I did it using PythonSDK), I could not upload the whole set (~9000 mel-spectrograms). When I would queue an upload, it would just break after 13 minutes. That is why I uploaded a smaller chunk of the testing set (but still, I need to see the performance of the quantized model on the full test set). Is it not possible to upload more data?
How do I quantize my model and see the test accuracy and calculated inference time for the deployed model?
Additionally, as I did not go through the processing block creation, I am wondering if there is any preprocessing code for the mel spectrograms available for the Arduino library that I can use to calculate mel-spectrograms on the audio collected from the onboard mic?
Thank you.
Kind regards