torchio
Use TorchIO 2.0 for medical image loading, preprocessing, augmentation, patch-based training/inference, CLI conversion, and I/O troubleshooting.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- torchio
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use TorchIO 2.0 for medical image loading, preprocessing, augmentation, patch-based training/inference, CLI conversion, and I/O troubleshooting.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-quick-map.md2.4 KB
- 参考文档references/repo-provenance.md1.2 KB
- 参考文档references/repo-routing-metadata.json409 B
- 参考文档references/troubleshooting.md2.9 KB
03 SKILL.md 正文
TorchIO
Use this skill when working with TorchIO, a PyTorch-oriented library for 3D medical image loading, preprocessing, augmentation, patch sampling, dense inference, and command-line image utilities.
Start Here
- Confirm the package imports:
import torchio as tio print(tio.__version__) - Use TorchIO 2.0 names in generated code:
Flip,Affine,Noise,Motion,BiasField,Spike, andGhostinginstead of oldRandom*class names. - Build in-memory images with
tio.ScalarImage(tensor)ortio.LabelMap(tensor)using 4D(C, I, J, K)tensors. - Keep runtime examples self-contained with synthetic tensors or user-provided image paths; do not depend on TorchIO repository examples or tests.
Route by Task
- Data containers and validation: use data-model for
Image,ScalarImage,LabelMap,Subject/Study, affines, points, bounding boxes, lazy/eager source handling, saving, and source-shape errors. - Preprocessing and augmentation: use transforms for transform families,
Compose,OneOf,SomeOf, include/exclude routing, per-instance batches, history/inverse behavior, and MONAI/Cornucopia adapters. - Patch training and dense inference: use patch-workflows for
GridSampler,UniformSampler,WeightedSampler,LabelSampler,Queue,SubjectsLoader,PatchLocation, andPatchAggregator. - CLI, conversion, visualization, and remote I/O: use cli-and-io for the
torchioconsole entry point,info,convert,transform,cache, plotting/animation, NIfTI-Zarr, cloud/remote extras, and cache/download caveats.
Shared References
- Read API quick map when choosing which TorchIO class or sub-skill owns a task.
- Read troubleshooting for install/import, optional dependency, API drift, tensor shape, CLI, and workflow-routing failures.
- Read repo provenance before refreshing this skill against a newer TorchIO checkout.
Safe Smoke Checks
Run the nearest bundled script after installing TorchIO and its base dependencies:
python sub-skills/data-model/scripts/data_model_smoke.py
python sub-skills/transforms/scripts/transform_history_smoke.py
python sub-skills/transforms/scripts/include_exclude_smoke.py
python sub-skills/patch-workflows/scripts/patch_workflow_smoke.py
python sub-skills/cli-and-io/scripts/torchio_cli_smoke.py --print-only
Use full CLI smoke only when local temporary file creation and NIfTI conversion are acceptable:
python sub-skills/cli-and-io/scripts/torchio_cli_smoke.py
Install and Extras Guidance
- Base install should provide
torch,nibabel,SimpleITK,numpy,tyro,fsspec[http], and TorchIO core APIs. - Optional extras are task-specific:
plotfor Matplotlib visualization,videofor GIF/MP4 animation,monaiandcornucopiafor transform adapters,zarrfor NIfTI-Zarr, and cloud extras fors3,azure, orgcsremote storage. - CPU is sufficient for the skill workflows unless the user's own model or pipeline requires CUDA.
Common Decisions
- Use
ScalarImagefor continuous intensity data andLabelMapfor masks/segmentations so spatial transforms use correct interpolation. - Use
Subjectto keep multiple images and metadata aligned through transforms and patch sampling. - Use
include/excludetransform arguments when only some image keys should change. - Use
GridSamplerplusPatchAggregatorfor dense inference; useQueueplus random samplers for stochastic patch training. - Use the CLI for one-off inspection/conversion and the Python API for pipelines that need precise parameter control, batching, or model integration.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).