ViT-Base INT8

Edge LinuxRaspberry Pi 5

INT8-quantized Vision Transformer for 1000-class image classification, using 16×16 image patches at 224×224 resolution. It is designed for efficient edge inference on ARM devices with XNNPACK acceleration, producing raw ImageNet logits from RGB inputs while preserving accuracy with a much smaller, faster, and lower-memory footprint than the FP32 baseline.View on Hugging Face

Accuracyvs base model

Top-181.5%↓0.02 pp

Top-596.06%0 pp

Latency

About this version

Original Model:: vit-base-patch16-224 by Google — Repository
Optimization/conversion:: Arm-Optimized version for execution on Arm-based platforms.
Converted/optimized by:: Arm
License:: The Original Model and the Optimized Model are subject to Apache-2.0.

This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

About this version

Original Model:: vit-base-patch16-224 by Google — Repository
Optimization/conversion:: Arm-Optimized version for execution on Arm-based platforms.
Converted/optimized by:: Arm
License:: The Original Model and the Optimized Model are subject to Apache-2.0.

Original Model and Documentation

Licenses and Third-Party Terms

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Metric	Resultsvs baseline	Change
End to end latency p50	378.14ms1101.03ms	−722.89 ms
End to end latency p90	379.86ms1102.09ms	−722.23 ms
End to end latency p99	388.31ms1104.82ms	−716.51 ms
Peak memory	110.05MB394.34MB	−284.29 MB
Average memory	107.98MB394.28MB	−286.30 MB
Time to first inference	378.77ms1128.78ms	−750.00 ms
Frames per second	2.64fps0.91fps	+1.73 fps
Accuracy - Top 1	81.50%81.52%	−0.02%
Accuracy - Top 5	96.06%96.06%	0.00%

Configuration	Baseline	Arm Optimized
Dataset	ImageNet-1k (val)	ImageNet-1k (val)
Sample count	50,000	50,000
Batch size	1	1
Input resolution	224x224	224x224
Runs	100	100
Warmup runs	10	10

ViT-Base INT8

ViT-Base INT8

Get Started with this Model

Image Classifier (ExecuTorch + XNNPACK)

Model Evaluation

Configuration

Comparability

Purpose of this Release