Custom DeepLabV3 + MobileNetV3-Large segmentation model HEF compilation error

General Discussion
1

Hi everyone,

I am trying to compile a custom DeepLabV3 + MobileNetV3-Large lane-segmentation model for Hailo-8 with DFC 3.33.1.

Current setup:

  • ONNX opset 11
  • static input: 1x3x384x384
  • 6 classes
  • custom exported ONNX from PyTorch
  • global avgpool ops already patched out
  • auxiliary head removed
  • conservative model script (optimization_level=0, compression_level=0, batch_size=1, calibset_size=64)

What happens:

  • hailo parser onnx succeeds
  • hailo optimize succeeds
  • partitioning succeeds
  • final compile fails during mapping with:

BackendAllocatorException: Compilation failed: Value doesn't fit in field (1411)

The hailo_sdk.core.log also shows multiple microcode violations around resize* / conv* layers before the failure.

I already tried reducing input size:

  • 512 → fail
  • 448 → fail
  • 384 → still fail

Question:
Is DeepLabV3 + MobileNetV3-Large / resize-heavy ASPP segmentation known to hit a compiler/backend limitation on Hailo-8?

Thanks.