quantized .pte model for image classification with Ethos-U55 or Cortex M CPU

[xianghui-renesas]

Hi, can you share any quantized .pte model for image classification with Ethos-U55 or Cortex M CPU? Appreciate your response!

[wilhe01]

I am not sure if you have seen the examples here: https://github.com/pytorch/executorch/tree/main/examples/arm/image_classification_example_ethos_u

[Sir-Sloth-The-Lazy]

The ExecuTorch repo has pre-quantized .pte examples in the examples/ directory.
For Ethos-U55 specifically, check:

examples/arm/
├── aot_arm_compiler.py # shows how to export + quantize for ARM backends
└── run.sh # end-to-end script
To generate a quantized MobileNetV2 .pte for Cortex-M / Ethos-U:

From the executorch root

python examples/arm/aot_arm_compiler.py
--model_name="mv2"
--delegate
--quantize
--target=ethos-u55-128
This produces mv2_delegate.pte which runs on Ethos-U55 via the Vela-compiled
delegate.

For a minimal classification example to verify the toolchain first, try MobileNetV2
at int8 before moving to custom models — the quantization flow for Ethos-U55 requires
XNNPACK-compatible ops, and custom layers often need manual annotation with
XNNPACKQuantizer.

The examples/arm/README.md also has links to pre-built .pte files for CI testing.
Does your environment have the Arm Compiler (armclang) and Vela installed?