The yolo-artifacts volume also keeps metrics and annotated images after both
containers exit.
The default run trains on Ultralytics’ eight-image COCO8 dataset for one
epoch. Use it to check training, checkpoint storage, and prediction before
spending GPU time on your own dataset.
app.py in the downloaded project defines both functions and their resources.
The CLI installs once with uv tool install lazycloud-client. The commands
below download a standalone project and install its dependencies, with no
repository checkout. The rest of the guide runs from that directory through
uv run, which uses the project’s pinned SDK.
You need access to an L4 and a positive credit balance. Training and prediction
both use GPU compute. Check Ultralytics licensing
before adapting the example for commercial use.
Define shared storage
Both functions use the pinned Ultralytics image in app.py. Training writes a
checkpoint and metrics, and prediction reads that checkpoint from the same
volume. The code rejects paths outside the volume and refuses to overwrite an
existing run directory.
Both decorators include volumes=[artifact_volume]. The first invocation
creates the volume if it does not exist, and later calls reuse it. Importing
the module creates no remote resources.
Torch, CUDA, and Ultralytics stay inside the image. Nothing installs on your
machine.
Train
With no arguments, the function uses coco8.yaml, YOLO26n, and one epoch:
The command shows training logs and returns checkpoint and metrics paths.
Download the metrics CSV:
LazyCloud charges compute while the container runs, including startup and its
idle window. For this one-epoch run, startup and the model download often take
longer than training. Check the Usage page before raising epochs, image size,
or dataset size.
Predict
The prediction function loads the saved best.pt and runs it on Ultralytics’
public bus.jpg by default:
Open bus-prediction.jpg to see the detected objects. A one-epoch run only
checks the workflow. Evaluate accuracy on your own data before using the
predictions in a product.
Use your own data
Upload the dataset and its Ultralytics YAML to the volume. Paths inside the
YAML must also point at files under /artifacts.
Create the volume and upload from Python. This works even if you have not run
the sample:
create() reuses an existing volume. After the upload, train on your data:
To predict on an uploaded image, pass its volume-relative path as the third
argument. The first two name the training run and the prediction run:
A run name works only once, so earlier checkpoints stay intact. Give each new
trial a new training name and pass that name to prediction.
If training fails, find the task with uv run lazycloud task list --app yolo_training
and read its logs. Check the dataset path before retrying with a new name.
Clean up
The function containers shut down after their idle window. Keep the volume
while you still need the checkpoints, and download the files you want before
deleting anything.
Remove runs you no longer need:
Or delete the whole volume:
Deleting the volume deletes every checkpoint in it. Download what you want
to keep first.
Licenses
- The image is Ultralytics
8.4.101 and the model is yolo26n.pt. Read the
Ultralytics licensing options before
distributing a model or using one commercially.
- COCO8 is the first
eight COCO 2017 images and labels. Check COCO’s terms and each image’s
license before reusing them outside this test.
- Your own data is your responsibility: collection rights, annotation
license, privacy, and permitted model use.
