Arm/yolov9s-int8-xnnpack-executorch

yolov9s optimized for Arm-based Edge Linux

An INT8-quantized version of yolov9s for object detection, exported to ExecuTorch (.pte) and optimized for Arm-based Edge Linux systems.

Summary

This repository contains an Arm-optimized version of yolov9s for object detection, quantized to INT8 via static post-training quantization — per-channel symmetric weights, per-tensor affine activations — applied to the backbone and neck, with the stem convolution and the complete detection head kept in FP32. The model is provided in ExecuTorch (.pte) format, targeting Edge Linux systems.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on COCO 2017 and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	ExecuTorch (.pte)
Target device class	Edge Linux
Reference device	Raspberry Pi 5 (Cortex-A76, Raspberry Pi OS 64-bit based on Debian 13 "Trixie")
Primary performance result	p50 latency 463.89 ms (1.28x faster than baseline), 2.16 FPS
Accuracy result	mAP@0.5:0.95 44.00%
Size / memory result	11.87 MB (2.35x smaller), peak memory 108.97 MB

Original model

Field	Value
Original model	yolov9s
Original source	GitHub
Original developer	YOLOv9 (Chien-Yao Wang and Hong-Yuan Mark Liao)
Original model card	YOLOv9 repository
Original license	GPL-3.0

Model files

File	Description
yolov9s_raspberry_executorch_optimized.pte	Arm-optimized model for deployment
example.py	Minimal inference example
pyproject.toml	Pinned runtime dependencies for example.py, resolved with uv
uv.lock	Locked dependency resolution for pyproject.toml
config.yaml	Model I/O contract used by the example
benchmarks/	FP32 baseline and Arm-optimized benchmark records

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	Raspberry Pi 5
CPU / accelerator	Cortex-A76, 4 cores @ 2.4 GHz (CPU execution)
OS	Raspberry Pi OS 64-bit, based on Debian 13 "Trixie"
Runtime	ExecuTorch 1.1.0
Backend / delegate	XNNPACK, KleidiAI
Batch size	1
Precision	INT8 static PTQ — per-channel symmetric weights, per-tensor affine activations, applied to the backbone and neck; the stem convolution (model.0.conv) and the complete detection head (model.22) remain FP32
Runs	10 warmup + 50 measured

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 latency (ms)	596.094	463.888	1.28x faster
p90 latency (ms)	597.267	465.001	1.28x faster
p99 latency (ms)	599.931	466.869	1.29x faster
Throughput (FPS)	1.68	2.16	1.29x
Model load time (ms)	39.792	25.988	1.53x faster
Time to first inference (ms)	648.434	487.94	1.33x faster
Model size (MB)	27.907	11.873	2.35x smaller
Peak memory (MB)	145.42	108.97	1.33x less
Average memory (MB)	143.42	106.45	1.35x less

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Field	Value
Dataset	COCO 2017
Split	val2017
Number of samples	5000
Metric(s)	mAP@0.5, mAP@0.75, mAP@0.5:0.95
Evaluation runtime	ExecuTorch 1.1.0

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
mAP@0.5 (%)	63.10	61.90	-1.20 pp
mAP@0.75 (%)	50.93	48.50	-2.43 pp
mAP@0.5:0.95 (%)	46.77	44.00	-2.77 pp

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

Optimization area	Applied?	Notes
Model conversion	Yes	Converted to ExecuTorch .pte
Quantization	Yes	INT8 static PTQ — per-channel symmetric weights, per-tensor affine activations, calibrated on 2,000 random COCO 2017 samples; the stem convolution (model.0.conv) and the complete detection head (model.22) are kept in FP32 to preserve accuracy and avoid non-delegated dequantize_per_channel operations
Runtime/backend selection	Yes	XNNPACK + KleidiAI delegate
Graph/runtime compatibility updates	Yes	Performed as part of the ExecuTorch export pipeline: residual Q/DQ pairs inserted by PT2E are removed from detection-head concatenations after conversion, preventing a shared per-tensor scale from compressing class scores in the final box/class-score merge
Accuracy validation	Yes	Compared against the original model or published baseline
Performance validation	Yes	Measured on the reference Arm platform

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

Install dependencies

Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:

uv python install
uv sync --frozen

Run the example

uv run example.py

Expected input

Property	Value
Input shape	[1, 3, 640, 640]
Input type	float32
Input range	[0.0, 1.0]
Preprocessing	letterbox-resize to 640x640, pad with a constant value of 114/255 (approximately 0.4471), convert to tensor; no normalization

Expected output

Property	Value
Output shape	[1, 84, 8400]
Output type	N/A
Postprocessing	transpose to [1, 8400, 84]; multi-label decode (each candidate/class pair above the confidence threshold of 0.001 becomes its own detection); convert boxes from xywh to xyxy; sort and cap at 30000 candidates; batched per-class NMS at IoU threshold 0.7, capped at 300 final detections

Intended use

This model is intended for developers evaluating object detection workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• Accuracy was evaluated on COCO 2017 val2017 and may not generalize to all domains.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.

Additional notes

The stem convolution (model.0.conv) and the complete detection head (model.22) were skipped from INT8 quantization and kept in FP32 to preserve accuracy. Letterbox preprocessing preserves aspect ratio via padding, but the fixed 640x640 input size still requires resizing, and other input resolutions would require re-export.

• Sample input: sample_input.jpg is derived from Living room (Unsplash) by Jarosław Ceborski, via Wikimedia Commons (CC0 1.0).

About this version

Original Model: yolov9s by YOLOv9 (Chien-Yao Wang and Hong-Yuan Mark Liao) - 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 GPL-3.0.

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.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.