YOLOv7 .pt to .nb Local conversion and deployement for AMB82-MINI – No Bounding Box or Score Output

Technical Discussion
,
1

Hello,

I’m working on deploying a custom-trained YOLOv7 model to the AMB82-MINI board (using Arduino SDK). I’ve been using the YOLOv7 repo by Wongkinyu and the online converter tool successfully. However, I’m now trying to convert the model offline using the instructions from Ameba’s NN Deployment Guide, and I’m running into issues.

My custom-trained model has 4 classes (n=4)

And here’s the pipeline I followed:

  • Training Config: cfg/yolov7-tiny-deploy.yamlfromameba-arduino-pro2 sdk

  • Reparameterization: Used reparam_yolov7-tiny.py → Output:
    No implicit layers found - weights may already be reparameterized

  • Converted to ONNX format (2 output layers: output0 and 1598)

    • output layer:output0
    • output layer: 1598

      • Used Acuity Toolkit for conversion from .pt to .nb:

        • Modified the Pegasus_import.sh script as follows:

        • For output0

        • function import_onnx_network()

            • {

              NAME=$1

              rm_json_data

              echo “=========== Converting $NAME ONNX model ===========”

              cmd="$PEGASUS import onnx\

              –model ${NAME}.onnx \

              –outputs ‘/model.29/Mul_output_0 /model.29/Sigmoid_output_0’ \

              –output-model ${NAME}.json \

              –output-data ${NAME}.data \

              $(cat inputs_outputs.txt)"

              }

            • Inference Result: Obtained “No Back Trace!” in Arduino Serial Monitor

          • For 1598

            • function import_onnx_network()

              {

              NAME=$1

              rm_json_data

              echo “=========== Converting $NAME ONNX model ===========”

              cmd="$PEGASUS import onnx\

              –model ${NAME}.onnx \

              –outputs ‘/model.29/Mul_1_output_0 /model.29/Sigmoid_1_output_0’ \

              –output-model ${NAME}.json \

              –output-data ${NAME}.data \

              $(cat inputs_outputs.txt)"

              }

            • Inference Result: YOLOV7 Loaded, but no Bounding Box or Score

        • Ran Pegasus Import, Quantize (uint8), and Export (received .nb)

      • Renamed the resulting .nb file to yolov7-tiny.nb and copied it to the SD card (enabled NN model load from SD Card in Arduino)

      • Ran inference on **AMB82MINI

        Problem:** During inference (for 1598), I’m not receiving any probability output or bounding box.

Question:
Has anyone successfully deployed YOLOv7 on AMB82-MINI using the offline pipeline? If you have access to a working example or additional documentation, I’d really appreciate it. Happy to share more details if needed.

Looking forward to your guidance.

BR,

Hari

2

Hi @harir

Can I check if you modify nc = 2 to nc = 4 when training your model?

image
image1024×337 13.6 KB