Request: Offline AI Model Conversion Tool Access for Traffic Sign Detection POC

Arduino AI
1

Hello,

I’m working on a traffic sign detection POC using AMB82-MINI and need access to the offline AI model conversion toolkit (Acuity Toolkit).

What I need:

  • Access to ameba-ai-offline-toolkit GitHub repository
  • Acuity conversion tools to convert TFLite/H5 model to .nb format

Current status:

  • Have AMB82-MINI board working
  • Camera and DEFAULT_IMGCLASS tested successfully
  • Have traffic sign detection model (TFLite format)
  • Need to convert to .nb format for deployment

What I tried:

  • Emailed AmebaAIoT@realtek.com with subject “Offline AI Model” - no response yet
  • Online conversion tool at amebaiot.com - not receiving email with converted model

Request:
Could someone from Realtek team please grant access or provide alternative download method for the offline conversion toolkit?

My use case: ADAS lite system for road signs detection

Thank you!

2

Hello,

We have replied to your email request for the offline toolkit. Regarding your online model conversion issue, may I know what model are you using currently? Would you mind to share your model with us so that we can help to verify the problem?

Thank you.

3

Hello Kevin,

Thank you for reaching out. I am currently working with a custom YOLOv7-tiny model (Pytorch) for road sign detection on the AmebaPro2 (AMB82-mini).

  1. ONNX Layer Names: When converting my YOLOv7-tiny .pt to .onnx for the offline tool, which specific output nodes should I target to avoid the ‘No Bounding Box’ error?"
  2. Widescreen Grid Math: “For a custom 576x320 resolution, does the offline toolkit automatically handle the non-square grid mapping, or do I need to manually adjust the Pegasus import script’s outputs?”
  3. Reparameterization Necessity: "Is the reparam_yolov7-tiny.py script required even for offline conversion, or does the toolkit handle reparameterization during the .nb export?
4

Hi @krish_tyger ,

  1. You might need to modify the models/yolo.py file in the YOLOv7 repo. Kindly refer to the guide NN Example (Yolov7_tiny) to modify the forward function within class Detect. Please take note of the part of Convert PyTorch to onnx. You may also refer to this related forum thread, YOLOv7-tiny .nb (NBG) model inference shows no bounding boxes on Amb82-mini
  2. No, you do not need to manually adjust the pegasus import script’s outputs, it’ll be automatically handled by the toolkit.
  3. Yes, reparameterization script is needed as the toolkit does not handle reparameterization process.

Thank you.