How to train and deploy an ultra-small YOLO model on a constrained device (OpenArtPlus v1.0, 2MB SRAM, 64MB SDRAM)?

[GaoPeng415]

Hello everyone,

I am trying to train a model for my smart car to recognize tennis balls and teddy bears—this is a requirement for the 21st Smart Car Competition. More details about the competition can be found here: link.

We have a camera module (OpenArtPlus v1.0) which has 2MB SRAM and 64MB SDRAM, and supports TensorFlow Lite. You can find its specifications here: NXP AIoT Cloud and Gitee.

I have tried using the following YOLO models: yolo26n.pt, yolov5nu.pt, yolov8n.pt, yolov3tinyu.pt to train my model and export to TFLite with int8=True. However, the smallest final model is 2624KB, which cannot run properly on the camera. Meanwhile, the example model provided by the OpenArtPlus repository is only 103KB, which surprised me—I cannot figure out how it is so small.

I am new to YOLO and only have a basic understanding. Could anyone kindly give me guidance or advice on how to train a model for recognizing tennis balls and teddy bears that can run on this constrained device? My ultimate goal is to use the model to locate objects and control the smart car to lift them.

Thank you very much for your help!

P.S. I’m not sure if asking this kind of help here aligns with the community guidelines. If this post is inappropriate, please let me know and I will remove it. If possible, I’d appreciate suggestions on where it would be more appropriate to ask for help. =)

[UltralyticsAssistant]

👋 Hello @GaoPeng415, thank you for your interest in Ultralytics 🚀! This is an automated response to help you get moving quickly, and an Ultralytics engineer will also assist soon 😊

We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question (your constrained-device + TFLite export scenario sounds like it), please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

With only 2MB SRAM, the blocker is usually runtime RAM (activations + TFLite buffers + postprocess/NMS) rather than the .tflite weight file size, so an INT8 export in the ~2–3MB range is expected for standard YOLO detectors and still may not run; that ~103KB “example model” is very likely a tiny classifier or a vendor-specific packed artifact (not a YOLO-style detector), so it’s not an apples-to-apples comparison. If you still want to push YOLO as far as possible, try exporting with a much smaller imgsz (this won’t shrink weights much, but it can drastically reduce peak RAM) and make sure you’re doing a calibrated INT8 export as shown in the TFLite export guide (also see model quantization notes):

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.export(format="tflite", int8=True, data="your_data.yaml", imgsz=320)

If you can paste the exact on-device failure log (text) and share what that 103KB model file actually is (format + how it’s loaded), we can better tell whether you’re hitting SRAM limits, unsupported ops, or a postprocessing constraint.