Skip to main content

Train cockpit (/p/<project>/train)

Submit a training job against the current export.

  • Class selection defaults to classes with enough data, with an "exclude classes without enough data" option.
  • Model size, augmentation preset (served list), GPU allocation, and a debounced preflight report — every check renders generically by name/severity/message, so a new backend check needs no frontend change.
  • Live progress and a log tail while a run is active; multi-size campaigns get an auto-promote-best + stop-when threshold.
  • Region dataset — when the backend serves a region profile, the dataset card offers a second, single-class dataset built from that profile's region boxes, named after the profile. The screenshots below come from a backend running the example license-plate profile, so it reads "Plates dataset"; with a different profile it takes that profile's name, and with no region profile it isn't shown.

Past runs​

  • Results — test-split evaluation (labelled by its actual split, "val" vs "test"), per-class table, the confusion matrix image (only when the backend serves its URL), an MLflow link (only when the backend serves a public MLflow URL), and full lineage (export identity, class remap, code versions), fetched lazily on first open.
  • Promote ↑ — to the Triton model registry, with the served gate report and, when the backend allows it, a "promote anyway" override.
  • Reproduce — resubmits the same spec/lineage as a fresh job.
  • Run probe predictions — scores the pool with the run's checkpoint, which is what feeds the Uncertainty and Model Disagreements review queues.
Train — preflight, live progress, past-run results
Finished-run Results — metrics, evaluation split and the confusion matrix the trainer wrote

Training cohorts​

A per-class cohort picker sourced from the backend (GET {API_PREFIX}/training_cohorts): 4 class-agnostic cohorts plus, with a region profile, 5 region-training-candidate modes. Thresholds shown are the server's own, never a client constant. This is a curation preview only — it never filters the actual training run.