Training

TL;DR

• You bring annotated images; Pipelogic runs a trainer on them for a bounded time and keeps the model file.
• A dataset has a task (detect, segment, classify, pose or obb) and versions. Only a validated, ready version can be trained on.
• A run is placed on a workspace node (unbilled) or a reserved cloud machine paid up front in PLC; unused time is refunded.
• Every epoch is reported while the run goes; the latest checkpoint is kept, so an interrupted run resumes instead of restarting.

Datasets

A dataset is a named collection of annotated images owned by the workspace. Its task is fixed when you create it:

Task	Annotations
detect	boxes with a class
segment	masks or polygons with a class
classify	one class per image
pose	keypoints per object
obb	rotated boxes with a class

Content lives in versions: each is one uploaded archive, validated for the dataset's task. A version is validating, then ready, or invalid with the reason. To change the data, add a new version.

A ready version reports images and annotations per split, instances per class, image sizes, unannotated images, issues, and up to 12 annotated previews. Check them before you train.

Layouts

Layout	Structure
COCO	train/, valid/, test/, each with its images and an _annotations.coco.json
YOLO	a data.yaml, the images and the label files
Classification	one folder per class, holding that class's images

Trainers

A trainer is a component in the category Train. It declares:

• a required dataset file slot (a bundle) and optional pretrained and resume slots;
• generated slots model, metadata and checkpoint or state;
• parameters epochs and max_train_minutes, plus its own.

metadata is JSON with format, task, class_names and metrics.final.

While it runs, a trainer writes two kinds of line to its standard output:

PPL_TRAIN_EPOCH {"epoch": 1, "train/loss": 5.993, "val/mAP_50": 0.033, "val/mAP_50_95": 0.028}
PPL_TRAIN_CHECKPOINT {"epoch": 1, "file": "state-epoch1.ckpt"}

The first reports an epoch and its metrics; the second names a checkpoint just written. The platform keeps the latest.

Runs

• Duration. From 30 minutes to 168 hours. The last 15 minutes are reserved for exporting the model.
• Placement. A workspace node (unbilled) or a cloud machine reserved for the run. By default, the cheapest that fits.
• Price. Paid up front in PLC, with the unused part refunded. An estimate shows the price before anything starts.
• Approval. When an agent starts a run, the person approves each run at its price.
• Interruptions. A run that loses its machine resumes from its latest checkpoint. An ended run can be resumed, and a running one canceled; the checkpoint is kept and the unused reservation refunded.

Related

• Train a model - the step-by-step flow.
• Model artifact integration - serve the trained model in a backend.
• Models - how model files reach a runtime.