Train a model

TL;DR

• Create a dataset, add a version of annotated images, wait until it is ready.
• Pick a trainer from the Train category, estimate the price, start the run.
• Follow the epochs while it runs; take the model, metadata and checkpoint files when it ends.
• Hand the model file to a backend with Model artifact integration.

Steps

1. Create the dataset with a name and a task: detect, segment, classify, pose or obb.

ppl dataset create --name street-signs --task detectCopy

2. Add a version from a folder, a .tar archive or an uploaded archive. Layouts are COCO (train/, valid/, test/ with _annotations.coco.json), YOLO (data.yaml, images, labels) or one folder per class.

ppl dataset version add <dataset_id> ./street-signsCopy

3. Wait for ready. Check the images per split, instances per class, unannotated images and issues. An invalid version shows why.

ppl dataset version get <dataset_id> 1Copy

4. Choose a trainer.

ppl component list --category TrainCopyppl component get <trainer>Copyppl component versions <trainer>Copy

5. Estimate the price. The estimate shows where the run is placed and the PLC it commits.

ppl train estimate --dataset-version <dataset_version_id> --version <trainer_version_id> --max-duration 4h --name street-signs-v1Copy

6. Start the run. Same flags. The maximum duration is 30 minutes to 168 hours; the last 15 minutes are reserved for exporting the model. A cloud run is paid up front and unused time is refunded.

ppl train start --dataset-version <dataset_version_id> --version <trainer_version_id> --max-duration 4h --name street-signs-v1Copy

7. Follow it and take the result.

ppl train get <run_id> --followCopyppl train get <run_id>Copy

The result lists the final metrics, the PLC committed and spent, and the output files by slot.

Resume or cancel

train resume <run_id> continues from the latest checkpoint of an ended run. train cancel <run_id> stops a run, keeps its checkpoint and refunds the unused reservation.

Related

• Training - datasets, the trainer contract, runs.
• Model artifact integration - serve the trained model.
• File upload and binding - bind files to components.