Classify pet images with DeiT-Tiny and Arm VGF using ExecuTorch
Introduction
Understand the DeiT-Tiny deployment workflow
Prepare ExecuTorch and build the VGF runner
Fine-tune DeiT-Tiny on pet images
Quantize and export DeiT-Tiny to VGF
Classify a pet image and verify VGF execution
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

About this Learning Path

Skill level:  Advanced
Reading time:  2 hrs
Last updated:  02 Oct 2026
Skill level:
  Advanced
Reading time:
  2 hrs
Last updated:
  02 Oct 2026
Author:Usamah Zaheer, Arm
Arm IP:Mali
Tags: ML Linux ExecuTorch PyTorch Python VGF Vulkan CMake Hugging Face
Author:
  • Usamah Zaheer, Arm
Arm IP:
Mali
Tags:
ML Linux ExecuTorch PyTorch Python VGF Vulkan CMake Hugging Face

Who is this for?

This Learning Path is for machine learning (ML) developers who want to deploy an image classifier with ExecuTorch and Arm VGF using the ML SDK for Vulkan's host emulation layers.

What will you learn?

Upon completion of this Learning Path, you will be able to:

  • Prepare ExecuTorch, the ML SDK for Vulkan, and a VGF runner on a Linux host.
  • Fine-tune DeiT-Tiny on the Oxford-IIIT Pet dataset and export a quantized VGF-backed ExecuTorch program.
  • Classify a pet image with the VGF-backed ExecuTorch program.
  • Compare the predicted breed with the dataset label and confirm VGF execution on the host.

Prerequisites

Before starting, you will need the following:

  • An Arm Linux development machine
  • A working Vulkan 1.3 or later GPU driver with shaderFloat64 support for the packaged ML emulation layer
  • Internet access and sufficient disk space for the source code, SDK, Oxford-IIIT Pet dataset, and model checkpoints

Summary

AI-assisted

This summary was drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.

Close
?
You’ll fine-tune DeiT-Tiny for pet breed classification, export it as a quantized VGF-backed ExecuTorch program, and run it on an Arm Linux host. First, you’ll prepare ExecuTorch and build the VGF runner. You’ll then train the model on the Oxford-IIIT Pet dataset and prepare its checkpoint for export. Finally, you’ll classify a test image, decode the predicted breed, and confirm VGF execution.

Frequently asked questions

AI-assisted

These FAQs were drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.

Close
?
Which ExecuTorch release should I use?
Use the ExecuTorch 1.5.1 release for both the Python package and the native source. This release uses stable PyTorch and TorchAO.
Do the Vulkan packages install my GPU's driver?
No. The listed packages provide Vulkan libraries and tools but don’t install a vendor-specific driver. Ensure that your GPU’s Vulkan driver is installed and working before building and running the VGF runner.
What does the helper script do?
The Learning Path helper handles checkpoint compatibility, image preparation, and breed decoding.
How can I confirm that the export completed successfully?
Check that the exporter reports the output path in arm_test/deit_vgf/export.log. Then, run test -s arm_test/deit_vgf/deit_quantized_vgf.pte to confirm that the exported program exists and isn’t empty. The log also reports the host accuracy check on 100 test images.
What should I see after preparing the input image?
You’ll see the input image path, the tensor shape [1, 3, 224, 224], and the dataset’s Expected breed. The helper also saves input.bin for the runner. The Expected breed at this stage is a reference label, not a model prediction.
Next