---

name: training-evaluation description: "Train, evaluate, checkpoint, and troubleshoot Detectron2 models with safe commands and APIs." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Detectron2 Training & Evaluation

Use this sub-skill when the task is to run or prepare Detectron2 training/evaluation, customize trainers/hooks/evaluators/checkpointing, or diagnose training/eval failures. It assumes datasets and configs have already been selected or registered by the dataset/configuration sub-skills.

Fast Routing

• Build safe training/eval commands with scripts/train_command_builder.py; it prints commands only and never starts training.
• Plan evaluator/checkpoint setup with scripts/evaluation_plan.py; it prints checks and next steps only and never runs evaluation.
• Read references/training-workflows.md for standard Detectron2 train/eval driver semantics, DefaultTrainer, SimpleTrainer, hooks, launch, solver scaling, and resume/eval-only behavior.
• Read references/evaluation-checkpointing.md for evaluator selection, inference_on_dataset, verify_results, DetectionCheckpointer, and checkpoint loading/resume behavior.
• Read references/troubleshooting.md before starting long jobs, changing GPU count, using a custom dataset, switching Yacs/LazyConfig syntax, or debugging failed evaluation.

Safe Defaults

• Prefer a dry-run command/planning step before running a long job; stop for user approval when training would be long, multi-GPU, dataset-heavy, or hardware-dependent.
• Use a project-local driver with the standard Detectron2 parser shape: python PROJECT_TRAIN_DRIVER.py --config-file CONFIG --num-gpus N [--resume|--eval-only] KEY VALUE ....
• Use LazyConfig syntax only for Python config files: python PROJECT_LAZY_TRAIN_DRIVER.py --config-file CONFIG.py train.output_dir=... dataloader.train.total_batch_size=....
• Do not use the repository demo CLI as a runtime dependency; in this checkout its demo imports a missing vision.fair.detectron2.demo.predictor module.

Common APIs

• Yacs training: subclass DefaultTrainer, override build_evaluator, optionally override build_train_loader, build_optimizer, build_lr_scheduler, or register hooks.
• Minimal loop: use SimpleTrainer(model, data_loader, optimizer) when DefaultTrainer assumptions are too restrictive.
• Distributed entry: call launch(main_func, num_gpus_per_machine, num_machines=1, machine_rank=0, dist_url=None, args=()).
• Evaluation: use COCOEvaluator, SemSegEvaluator, DatasetEvaluator, DatasetEvaluators, and inference_on_dataset(model, data_loader, evaluator).
• Checkpointing: use DetectionCheckpointer(model, save_dir, trainer=trainer) and distinguish load, resume_or_load, --resume, and --eval-only.