AMB82-MINI on Edge Impulse

[Arif]

Has anyone ever know how to use AMB82-MINI on Edge Impulse?

[M-ichae-l]

Hi, thanks for trying ameba for your projects.

Edge Impulse doe support 3rd party embedded boards that compatible with Arduino.

Let us follow the steps to make use of the edge impulse.

1, we can make use of edge impulse to train the model we wanted. Then export the model.

• note that the models need to convert before used by ameba, lets try study how to process the using-customized-nn-model first
• At edge impulse, refer the following to export the model that have trained. In my case, I have trained mobilenetv2 for image classification. I prefer to use tensorflow lite file that supported by AMB82-mini lately. We just need to find the .tflite file.
• Remember to check the model classes order, in my case is Cups, Keys, unknown
• Also remember to download at least 1 image file from “image“ for future quantization, my case is “001.jpg“.
• 
  
  

2, refer to the guide to convert NN model, it can be done both online/offline

• Amebapro2 AI Convert Model – Realtek IoT/Wi-Fi MCU Solutions

• Realtek AmebaPro2 Model Conversion Tool — AmebaPro2's Documentation v0.1 documentation

• I am using the offline tool by docker setup at ubuntu env. Acuity_installation_docker

  • Follow the guide, remember to extract the 2 .tgz files downloaded
  • 
  • go to acuity_examples_c901149/Script and update pegasus_export_ovx.sh
  • 
  • 
  • create a folder with any name, mine is “tflite_m“, put the .tflite fileinside and rename it same as folder “tflite_m.tflite“, create a dataset.txt file Input the image name inside the dataset.txt file.
  • 
  • 
  • Back to Docker/Linux at terminal run
    • docker login ghcr.io -u
    • source install.sh
    • source run.sh
  • go to acuity_examples_c901149/Scriptrun
    • ./pegasus_import.sh tflite_m
    • go to tflite_m/ update tflite_m_inputmeta.yml
      • add_preproc_node: true
      • preproc_type: IMAGE_RGB888_PLANAR
    • remember the preproc_image_size, my case, 96 and 96
    • 
  • run
    • ./pegasus_quantize.sh tflite_m uint8
    • ./pegasus_export_ovx.sh tflite_m uint8
  • go to wksp/tflite_m_uint8_nbg_unify/
    • network_binary.nb is our converted model file

3, use Arduino IDE to upload the model into board (recommended to use SD card)

• some useful example guides

• I tried this example Image Classification, refer to following for my setups

  • USE_MODEL_META_DATA_EN set to 0
  • select CUSTOMIZED_IMGCLASS_MOBILENETV2
  • change NNWIDTH and NNHEIGHT as 96
  • Update the class at ClassificationClass_list.h, in my case is Cups, Keys, unknown, use the same order get from the edge impulse export files.
  • 
  • select the SD card NN model mode at Arduino IDE
  • 
  • For SD card, create a folder NN_MDL , put the network_binary.nb file inside, and rename it as mobilenetv2_int16.nb
  • 

  

  • compile and upload firmware to AMB82-mini, refer to the guide for running and testing the NN application.