lightning-training
Build fastMRI PyTorch Lightning training and test workflows with FastMriDataModule, UnetModule, VarNetModule, masks, transforms, checkpoints, and CPU-safe debug commands.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- lightning-training
- license
- MIT
- metadata
- { "disco-role": "operating" }
- 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
02 资源文件 3
- 参考文档references/api-reference.md8.3 KB
- 参考文档references/training-workflows.md7.7 KB
- 参考文档references/troubleshooting.md7.9 KB
03 SKILL.md 正文
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_valruns.
Routing Rules
- Prefer
UnetModulewithUnetDataTransformfor baseline image-domain training on singlecoil or multicoil challenge data. - Prefer
VarNetModulewithVarNetDataTransformfor multicoil end-to-end variational network training; keep VarNet challenge asmulticoilunless the surrounding code has explicitly implemented a different path. - Use
create_mask_for_mask_typewithrandomfor the classic knee U-Net demo andequispaced_fractionfor brain/multicoil-style and VarNet workflows; exact mask classes includeEquiSpacedMaskFuncandEquispacedMaskFractionFunc. - For DDP, set the data module
distributed_sampler=Trueso validation/test volumes dispatch coherently, and let worker seeding handlemask_funcrandomness per worker/rank. - For CPU-only or quick verification, render a one-device
fast_dev_runor one-batch debug path instead of copying demo GPU/DDP defaults. - For leaderboard-style training, use
combine_train_val=Trueonly after validation choices are intentional, then test with explicittest_pathortest_splitand save reconstructions under the Lightningdefault_root_dir.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.