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?

[usamahz]

Hi @xianghui-renesas, MobileNetV2 is a good starting point. The repository provides export flows to generate a quantized .pte for either Ethos-U55 or a Cortex-M CPU.

After completing the Arm environment setup, the current main branch’s export script supports the following commands from the repository root. These have been checked against the source, but have not been execution-tested.

# Ethos-U55 with 128 MACs
python -m backends.arm.scripts.aot_arm_compiler \
  --model_name=mv2 \
  --delegate \
  --quantize \
  --target=ethos-u55-128 \
  --output=mv2_ethos_u55.pte

# Cortex-M55 CPU using CMSIS-NN
python -m backends.arm.scripts.aot_arm_compiler \
  --model_name=mv2 \
  --quantize \
  --target=cortex-m55 \
  --output=mv2_cortex_m55.pte

For classification accuracy, supply representative, preprocessed calibration inputs through --calibration_data; otherwise the script uses random example inputs. The Cortex-M MobileNetV2 notebook walks through dataset preparation, calibration, and execution.

Two clarifications to the earlier replies:

• The linked DeiT image-classification example currently targets Ethos-U85-256.
• Ethos-U uses EthosUQuantizer; XNNPACK compatibility is not a requirement for this flow.

Which board/SoC and ExecuTorch version are you using? For an Ethos-U55 .pte, the accelerator and memory settings should match your target configuration.