Fine-tune a vision model on your images

Train a detector for your own objects with Ultralytics YOLO. Put your images and labels in YOLO format into a folder, one text file per image with one line per object: class, then the box center, width and height as fractions of the image size.

vision/
  data.yaml
  train.py
  shapes/
    images/train/0001.jpg …   labels/train/0001.txt …
    images/val/0001.jpg …     labels/val/0001.txt …

data.yaml names the folders and the classes:

# Paths are relative to this file.
train: shapes/images/train
val: shapes/images/val
names:
  0: circle
  1: square

train.py starts from pretrained weights, trains, evaluates and exports the model to ONNX:

from pathlib import Path
from ultralytics import YOLO

out = Path("output").resolve()
model = YOLO("yolo11n.pt")  # pretrained weights, downloaded on first use
model.train(data="data.yaml", epochs=30, imgsz=320, batch=32, project=out, name="shapes", exist_ok=True)
metrics = model.val()
print(f"mAP50-95: {metrics.box.map:.3f}")
model.export(format="onnx")  # next to best.pt: output/shapes/weights/best.onnx

Run it:

nido create train --gpu l4 --network github.com --network release-assets.githubusercontent.com \
  --network pypi.org --network files.pythonhosted.org
nido cp ./vision train:/sandbox/vision
nido exec train -- uv pip install onnx onnxslim onnxruntime
nido exec train -- bash -c "cd /sandbox/vision && python train.py"
nido cp train:/sandbox/vision/output ./output

On 160 small images this takes about a minute. output/shapes/weights then holds best.pt and best.onnx, next to the training curves and a confusion matrix. The weights are yours.

To keep the run going in the background, or to run it as one task, see Long runs.