Image Classifier (ExecuTorch + XNNPACK)

An on-device evaluation harness for compatible ExecuTorch image classification models that use the XNNPACK backend and are hosted on Hugging Face. Deploy the Project to an Arm Target, upload images through a Gradio web interface, and measure inference results and performance. No GPU is required.

View on GitHub

Image Classifier (ExecuTorch + XNNPACK)

This project is a Topo Project and follows the Topo Project Specification.

This Project provides an on-device evaluation harness for Arm-optimized ExecuTorch image classification models that use the XNNPACK backend. Select a compatible .pte model, deploy the Project to an Arm Target, and upload images through the web interface. The dashboard reports prediction results and performance for the Target.

The Project demonstrates:

A multi-stage Docker build that downloads a Hugging Face model and embeds it in the image. The Target does not need the Hugging Face token or network access.
A configuration-driven CPU inference runner that reads configuration from the selected Hugging Face model repository.
Per-image latency benchmarking that reports the median and p90 after warmup runs.

Model compatibility

This Project supports only ExecuTorch image classification models that use the XNNPACK backend and are hosted on Hugging Face. Each model repository must contain config.yaml and metadata.yaml. These files identify the .pte file and configure preprocessing and postprocessing. A repository can also provide human-readable output labels in imagenet_classes.json. If it does not and metadata.yaml identifies calibration.dataset_name as ImageNet-1K, the runtime uses the categories from Torchvision's SqueezeNet1_1_Weights.IMAGENET1K_V1 metadata. Other models without labels use generic class indices.

Build-time parameters

The HF_REPO_ID and HF_ENDPOINT Project parameters are passed to Docker as build arguments and resolved at build time.

Parameter	Required	Description	Default
`HF_REPO_ID`	No	ExecuTorch + XNNPACK model repository on Hugging Face	`Arm/vit-base-int8-xnnpack-executorch`
`HF_ENDPOINT`	No	Hugging Face API endpoint	`https://huggingface.co`

Usage

Install Topo by following the instructions in the Topo repository.

Clone the Project

The clone step will prompt you for values for the HF_REPO_ID and HF_ENDPOINT parameters. Leave either input empty to select its default.

topo clone https://github.com/Arm-Examples/topo-executorch-image-classifier.git

Build and deploy the Project

Set a Hugging Face read token on the Host, and deploy the Project to the Target:

cd topo-executorch-image-classifier
topo deploy --target <user@hostname>

Topo builds the image on the Host and transfers the finished image to the Target over SSH. The Target does not need network access to download the model.

Note: To download a private model at build time, set HF_TOKEN on the Host before running topo deploy. The token must have read access to the repository. The build mounts it as a secret and does not store it in the image or transfer it to the Target. Public repositories do not require a token.

Open the web interface

After deployment, open http://<target-ip>:7860 in a browser. Upload an image to see the top predicted classes, confidence scores, and inference latency for the Target.

Run this project

Topo project

Use Topo to configure and deploy this container-based application to Arm-based Linux devices over SSH.

Configure project

Model

Deploy project

Install the Arm Topo CLI to manage your model deployments through the terminal

1. Install Topo CLI

2. Run the following command to clone and configure the project.

Clone and Configure Project

topo clone https://github.com/Arm-Examples/topo-executorch-image-classifier.git#e054e67367a7bb729983c8ec81ee16b58ca30cd3 HF_ENDPOINT='https://ml-models.devplatform.arm.com/' HF_REPO_ID='Arm/swin-tiny-int8-xnnpack-executorch'

Read the Topo docs for installation, troubleshooting and advanced usage.

Loading example…

About the models on this Arm AI Portal

The Arm AI Portal (the “Portal”) provides details of AI models developed, optimized or tested for use on Arm-based platforms. Models listed on the Portal may have been developed by Arm or by third parties. Unless expressly stated otherwise, Arm did not develop or train the underlying third-party model and is not responsible for its original intended behavior or use.

Information accompanies each model listing and describes the model’s provenance, together with relevant information about the original third-party model, its developer and repository, where applicable. Inclusion on the Portal of a model developed by a third party or based on a third-party model does not constitute Arm’s endorsement or certification of that third-party model or any third-party work relating to it.

Except as expressly stated in the information accompanying a specific model, Arm makes no representation as to the accuracy, safety, security, non-infringement, legal compliance, suitability for production use or fitness for any particular purpose of any model, converted or optimized version, or related outputs.

Inclusion of a model on the Portal does not itself grant you any rights to use that model. Use of each model is subject to the applicable license terms, usage restrictions and documentation. You are responsible for reviewing that information and independently evaluating the model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements. Any use of or reliance on a model, its outputs or related materials is at your own risk. Where the Portal links to repositories or other materials controlled by third parties, those repositories and materials are outside Arm’s control and may change without notice.

Use of the Portal is subject to the Arm Website Terms and Conditions of Use.