---

name: lightning-training description: "Build fastMRI PyTorch Lightning training and test workflows with FastMriDataModule, UnetModule, VarNetModule, masks, transforms, checkpoints, and CPU-safe debug commands." disable-model-invocation: true metadata: disco-role: operating license: MIT

fastMRI Lightning Training

Use this sub-skill when an agent needs to train, debug, resume, or test fastMRI UnetModule or VarNetModule workflows with PyTorch Lightning. Keep raw HDF5 layout and split details in data-loading, raw model constructor details in model-architectures, and final metric/submission packaging in evaluation-submission.

Start Here

• Use references/training-workflows.md when writing an end-to-end train/test script, adapting a U-Net or VarNet demo, choosing CPU fast-dev settings, or preserving DDP volume dispatch.
• Use references/api-reference.md when wiring exact FastMriDataModule, UnetModule, VarNetModule, mask, transform, checkpoint, or test-output APIs.
• Use references/troubleshooting.md when diagnosing Lightning version drift, missing directory config, CPU-only runs, mask/challenge mismatches, sample-rate conflicts, cache/import failures, checkpoint resume, or reconstruction paths.
• Use scripts/build_training_command.py to render explicit, source-independent command guidance for U-Net or VarNet training, test, CPU debug, DDP, and leaderboard-style combine_train_val runs.

Routing Rules

• Prefer UnetModule with UnetDataTransform for baseline image-domain training on singlecoil or multicoil challenge data.
• Prefer VarNetModule with VarNetDataTransform for multicoil end-to-end variational network training; keep VarNet challenge as multicoil unless the surrounding code has explicitly implemented a different path.
• Use create_mask_for_mask_type with random for the classic knee U-Net demo and equispaced_fraction for brain/multicoil-style and VarNet workflows; exact mask classes include EquiSpacedMaskFunc and EquispacedMaskFractionFunc.
• For DDP, set the data module distributed_sampler=True so validation/test volumes dispatch coherently, and let worker seeding handle mask_func randomness per worker/rank.
• For CPU-only or quick verification, render a one-device fast_dev_run or one-batch debug path instead of copying demo GPU/DDP defaults.
• For leaderboard-style training, use combine_train_val=True only after validation choices are intentional, then test with explicit test_path or test_split and save reconstructions under the Lightning default_root_dir.