NanoTrack Head ONNX Export → Hailo Parser & Quantization Errors (Unexpected zero dimension, Conv3D integer division)

[Mehmet_Demir]

1. Model: NanoTrack (Siamese tracker, MobileNetV3 backbone, BAN head)
2. Head type: DepthwiseBAN (ban_v2.py)
3. Checkpoint: nanotrackv2.pth (48-channel version)
4. Config: BAN.TYPE = DepthwiseBAN (configv2.yaml)
5. Exported with: PyTorch 1.13.1 → ONNX opset 14
6. Target HW: Hailo-8L
7. SDK Versions:

   • Hailo DFC 3.29.0

   • HailoRT 4.19.0

     I’m trying to run the NanoTrack head (DepthwiseBAN, MobileNetV3 backbone) on Hailo-8L.

     • ONNX export works fine.

     • Hailo parser runs successfully and produces a .har.

     • The problem appears during optimization/quantization with calibration data.

     Errors I see:

     1. "float32 is not a valid CalibrationDataType"

     2. integer division or modulo by zero inside Conv3DOp, log snippet:

        kernel shape: (1, 2, 2, 64), input feature: 16, groups: 64
        kernel shape 3: 64
        Exception encountered when calling layer 'conv_op' (type Conv3DOp).
        integer division or modulo by zero
        
        

     It looks like some depthwise conv layers in the BAN head are being mapped to Conv3DOp and break during quantization.

     Questions:

     • Is this a known limitation or bug in Hailo SDK with grouped/depthwise conv?

     • How should I fix the CalibrationDataType issue (is a different format expected)?

     • Is there a recommended way to restructure these layers so the head becomes Hailo-compatible?

     Thanks for any guidance!