Cloud Training

Link to this sectionLaunch a cloud training run#

In the Platform UI: Train → New Run. Pick:

The defaults inherit from the train mode reference — the same flags you'd pass to model.train(...) locally. When you start fine-tuning a previous checkpoint, the transfer learning glossary entry covers the why behind the lower LR; for systematic sweeps, the hyperparameter tuning guide is the canonical reference.

Link to this sectionPick the GPU thoughtfully#

The Cloud Training dialog exposes the full range — see the cloud training docs for the live list. A useful mental model, with rough hourly cost and VRAM:

For most academy-scale projects the default RTX PRO 6000 is the right call. As a reference point, an H100 SXM trains roughly 3.4× faster than an RTX 4090 — useful when you're paying for engineer time more than GPU time. Cost scales linearly with selected GPU rate × actual training hours; failed runs aren't billed. Higher tiers are also where distributed training and mixed precision start to pay back the configuration overhead.

Link to this sectionWatch the run live#

Every run has a live dashboard:

• Training/val loss — should both descend.
• Per-class AP — updates per epoch.
• Sample predictions — visually verify the model is learning what you think.
• Resource usage — GPU utilization should be > 80%; lower is wasted money.

Platform alerts you if a run looks unhealthy: loss diverging, GPU starved (input pipeline bottleneck), or out-of-memory.

Link to this sectionPulling the model locally#

Every Platform run produces a downloadable best.pt. The file is identical to what you'd produce locally with the same config — download it from the model page, then load it with the open-source package:

from ultralytics import YOLO

model = YOLO("best.pt")     # path to the file you downloaded
model.export(format="onnx")

You can now run inference, validate, and export exactly as in course 2. If you'd rather train locally and just stream metrics into Platform, set ULTRALYTICS_API_KEY and pass project=username/my-project name=experiment-1 to model.train(...) — see remote training.

Link to this sectionFine-tune from a previous run#

When you want to keep iterating, pick a previous model as the base model in the New Model dialog. Platform initializes the new run from your last best.pt automatically — drop the learning rate (e.g. lr0=0.001) in Advanced Settings for fine-tuning. The fine-tuning guide covers layer freezing and two-stage training when defaults aren't enough; augmentation knobs in Advanced Settings map directly to the data augmentation guide. The Cloud Training dialog also accepts a YAML or JSON config copied from any previous run, so you can reproduce or tweak earlier experiments without re-typing every parameter.

Link to this sectionWhat a healthy run looks like#

After ~30 epochs of training a small model on a clean dataset, you should see:

• Train and val loss both descending (not diverging).
• Val mAP@0.5:0.95 climbing past 0.4.
• Per-class AP not too unbalanced.
• Sample predictions visually correct.

If any of those don't hold by epoch 30, the dataset usually needs work — not the hyperparameters. The next lesson on experiment tracking gives you the tools to compare runs and find the dataset change that fixed things, and the Comet integration shows how to mirror Platform metrics into a third-party dashboard if your team already lives there.