Train a YOLO Detection Model on a Computer

This guide continues with the apple example. You will train YOLO11 on a computer, get best.pt, and export best.onnx for the board converter.

For the easiest first project, use online training on MaixHub. Computer training requires Python, but it keeps your data on your computer and gives you control over training.

Prepare the dataset first

Follow Prepare a Model and Dataset to arrange the pictures, labels, and data.yaml. The standard folder layout is explained there, so this page starts with installing the training tool.

Install the training tool

YOLO is a family of object-detection models. This guide uses the training tool provided by Ultralytics.

Install Python 3 from the Python website; Python 3.11 is recommended. On Windows, select Add Python to PATH during installation so the terminal can find it.

Open PowerShell on Windows, or Terminal on macOS or Linux. Check Python:

python --version

If you see Python 3..., install the training tool. If python is not found, use python3 in the commands instead.

python -m pip install ultralytics==8.4.104

Check the installation:

yolo checks

If the terminal prints the Python and environment information, training can start. If yolo is not found, close and reopen the terminal, then try again.

Run 3 rounds to check the data

This command trains for only three rounds. Its purpose is to find bad paths, images, or labels quickly:

yolo detect train model=yolo11n.pt data=/full/path/to/apple_dataset/data.yaml epochs=3 imgsz=640 project=runs/apple name=check exist_ok=True

yolo11n.pt is a small starter model. epochs=3 means the tool reads the training set three times. imgsz=640 means it processes the pictures at 640 pixels for training. Keep these values for the first check.

If the command finishes and creates runs/apple/check, the dataset format is usable. The first run downloads yolo11n.pt, so the computer must be online.

Train the model

Now train for 100 rounds:

yolo detect train model=yolo11n.pt data=/full/path/to/apple_dataset/data.yaml epochs=100 imgsz=640 project=runs/apple name=model exist_ok=True

When training ends, use:

runs/apple/model/weights/best.pt

best.pt comes from the round with the best validation result, which may not be the final round.

Test with a new picture

Choose an apple picture that was not used for training:

yolo detect predict model=runs/apple/model/weights/best.pt source=/path/to/test.jpg save=True

The terminal prints where it saved the result. Open that image and check the label and box.

If training pictures work but new ones do not, add pictures with different apples, backgrounds, and lighting. More training rounds usually cannot replace missing variety.

Export ONNX

The board converter cannot use .pt directly, so export it as .onnx.

Compatibility note for this example

This YOLO11 route uses ultralytics==8.4.104. Keep opset=17 and dynamic=False in the commands below. Other models or conversion routes may require different versions and settings; follow their own documentation instead of copying these values.

For MaixCAM and MaixCAM Pro, a common size is 320x224:

yolo export model=runs/apple/model/weights/best.pt format=onnx imgsz=224,320 opset=17 simplify=True dynamic=False

For MaixCAM2, start with 320x240:

yolo export model=runs/apple/model/weights/best.pt format=onnx imgsz=240,320 opset=17 simplify=True dynamic=False

For smaller objects, MaixCAM2 can use 640x480, but it runs more slowly:

yolo export model=runs/apple/model/weights/best.pt format=onnx imgsz=480,640 opset=17 simplify=True dynamic=False

The size order in these commands is height,width. A successful export creates best.onnx next to best.pt.

opset is the ONNX rule version; keep it at 17 for this route. dynamic=False fixes the input size, and simplify=True simplifies the model structure. Keep all three settings.

Convert and run on the board

1. Use Convert a YOLO Model Online to convert best.onnx into a board model package.
2. Follow Upload the model and test it.

When the camera image shows the correct label and box, training, conversion, and board deployment are complete. You may then share it on MaixHub, but sharing is not required.

Other YOLO versions

To try YOLO26 or YOLOv8, replace the starter model with yolo26n.pt or yolov8n.pt. Check the documentation for that model before reusing version or export settings.

The old YOLOv5 from the ultralytics/yolov5 repository is different from the newer YOLOv5u and cannot use this simple web-conversion route. For an old project, use manual conversion for MaixCAM or MaixCAM2. YOLO11 is recommended for a new project.

Pose, segmentation, and oriented boxes use different annotation formats and are not supported by this online conversion route. See Body Keypoints, Semantic Segmentation, or Oriented Bounding Boxes.