datasets-and-configuration
Prepare, select, and validate Gaussian-SLAM RGB-D datasets and YAML configurations for a safe GPU run.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- datasets-and-configuration
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Prepare, select, and validate Gaussian-SLAM RGB-D datasets and YAML configurations for a safe GPU run.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/configuration.md4.9 KB
- 参考文档references/data-and-downloads.md3.8 KB
- 参考文档references/data-formats.md5.5 KB
- 参考文档references/troubleshooting.md4.1 KB
- 脚本scripts/validate_config.py25.3 KB
03 SKILL.md 正文
Datasets and configuration
Use this sub-skill when a Gaussian-SLAM run needs an RGB-D scene, a dataset
alias, a scene-specific YAML file, or a preflight check. It covers the data
contract implemented by src/entities/datasets.py and the recursive YAML
merge behavior in src/utils/io_utils.py. It does not own SLAM
optimization, tracking/mapping choices, or evaluation metrics.
Applicability and operating contract
-
A run must use one of the exact dataset aliases
replica,tum_rgbd,scan_net, orscannetpp. These are case-sensitive;scannet,ScanNet,tum-rgbd, andscannet_ppare not aliases accepted byget_dataset. -
Start from a dataset default config and use a per-scene config with
inherit_from. The effective configuration must contain the top-levelproject_name,dataset_name,checkpoint_path,use_wandb,frame_limit,seed,mapping,tracking,cam, anddatakeys. A per-scenedata.frame_limitis also accepted and is the value passed to the dataset loader when present (the supplied ScanNet++ scenes use this form). -
Run the bundled validator before importing CUDA-heavy runtime modules:
python skills/disco/gaussian-slam/sub-skills/datasets-and-configuration/scripts/validate_config.py \ configs/Replica/office0.yamlAdd
--require-data --path-base <repository-root>when the data must be present now rather than merely checking the YAML contract. Multiple explicit config paths may be supplied. The validator uses PyYAML and standard Python modules only; it does not import PyTorch, OpenCV, Open3D, or custom CUDA extensions. -
run_slam.pyaccepts a config path followed by optional--input_pathand--output_pathoverrides. Those overrides replace the values underdata; validate the final values, not only the YAML defaults.
Safe workflow
- Identify the dataset and exact scene. Confirm that the external dataset's license, download permissions, credentials, and storage budget permit local use. Do not guess a scene's camera model or silently substitute a dataset.
- Inspect the corresponding default and scene YAML. Confirm inheritance,
dataset alias,
data.input_path,data.output_path, camera values, depth scale, and split/frame-limit behavior. Check both the run-levelframe_limitand any loader-leveldata.frame_limitoverride. - Acquire or mount data using the reference-only guidance in data-and-downloads.md. Never execute a download merely as part of validation.
- Validate the effective config and available data. Resolve relative input and output paths against the intended working directory (the source CLI does not rewrite them). Require paired RGB/depth/pose records and the metadata needed by the selected loader.
- Treat warnings as a stop-and-review gate before a GPU run. In particular, investigate relative paths, dropped timestamp associations, changed intrinsics, missing split names, and a frame limit larger than the selected sequence.
- Only after the preflight passes, launch the repository command with the final explicit paths. Preserve the effective YAML with the run output for reproducibility.
For field-level rules and effective merge examples, see configuration.md. For on-disk records and image, depth, and pose conventions, see data-formats.md. For download prerequisites and licensing boundaries, see data-and-downloads.md. For failure symptoms and recovery, see troubleshooting.md.
Expected observations and checks
A successful preflight reports the resolved inheritance chain, effective
dataset_name, camera dimensions/intrinsics/depth scale, selected split (when
applicable), and either verified data counts or an explicit “data path not
present; YAML-only validation” warning. With --require-data, absent paths,
missing files, malformed records, unsupported aliases, count mismatches, and
unresolved inheritance are failures (exit status 1).
The validator is intentionally conservative: it does not modify YAML, create output directories, download data, initialize wandb, allocate a GPU, or import the Gaussian-SLAM package. A passing validator proves configuration and layout preconditions only; it is not evidence that CUDA extensions, tracking, mapping, or evaluation will succeed.
Difficult synthetic usability cases
Use these cases during integrated skill verification; keep fixtures and reports outside this runtime skill tree:
- Inheritance and path trap: create
nested/scene.yamlinheritingbase.yaml, run validation from a different current directory, and give the scene a relativedata.input_path. The validator must load the sibling base by path relative toscene.yaml, flag the data path as relative, and reject a second fixture whose base path forms an inheritance cycle. - Silent frame drop trap: create a TUM fixture with valid-looking RGB and
depth lists but one pose timestamp more than
0.08seconds from every RGB timestamp; create a ScanNet++ fixture whose selected train split names a frame absent fromtransforms_undistorted.json. Validation must fail with a dataset-specific association/metadata message without importing CUDA.
Linked resources
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).