training-evaluation
Build and troubleshoot UniAD distributed and SLURM train/eval runs for BEVFormer, stage1, and stage2 workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- training-evaluation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Build and troubleshoot UniAD distributed and SLURM train/eval runs for BEVFormer, stage1, and stage2 workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/checkpoints.md3.2 KB
- 参考文档references/runtime-and-gpu.md3.5 KB
- 参考文档references/train-eval-cli.md7.1 KB
- 参考文档references/troubleshooting.md4.8 KB
03 SKILL.md 正文
Training and Evaluation
Use this sub-skill when the task is about launching, resuming, or troubleshooting UniAD training or evaluation runs for BEVFormer, stage1 track/map, or stage2 end-to-end workflows.
Route here for
tools/train.pyandtools/test.pycommand questions.tools/uniad_dist_train.sh,tools/uniad_dist_eval.sh,tools/uniad_slurm_train.sh, andtools/uniad_slurm_eval.shlauncher behavior.- Checkpoint placement,
work_dirs/layout, logs, and stage smoke checks. --seed,--deterministic,--cfg-options,--resume-from, orload_fromusage.- GPU-count, memory, or distributed-launch issues during training or evaluation.
Route elsewhere for
- Dataset download, info-PKL generation, CAN bus/map layout, or motion-anchor preparation ->
data-preparation. - Config semantics, model-head ownership, or architecture changes ->
config-and-model-architecture. - Result PKLs, visualization, or log plots ->
visualization-and-results.
Operating rules
- Prefer command construction, path validation, and dry-run diagnosis before recommending an expensive run.
- Treat
ckpts/at the repo root as the default checkpoint landing zone; the stage configs use relativeload_frompaths that expect that layout. tools/test.pyis distributed-only in practice: the non-distributed branch asserts false, so evaluation must go throughtorchrun,srun, or the bundled command builder.- The bundled launchers cap one node at 8 GPUs per node. Fewer GPUs are allowed, but runtime increases and exact metrics can drift.
- Quote list or tuple
--cfg-optionsvalues so the shell preserves brackets and commas. - Use
--resume-fromonly for resuming optimizer or scheduler state; useload_fromor--cfg-options load_from=...for initialization checkpoints. - When the user asks for an exact metric match, remind them that the published stage1 eval numbers are reference targets and can vary slightly with GPU count or launch topology.
Read/run next
- Read
references/train-eval-cli.mdfor CLI surfaces, launcher flags, launcher defaults, and work-dir/log behavior. - Read
references/runtime-and-gpu.mdfor the public v2.0 stack, memory guidance, and inspection caveats. - Read
references/checkpoints.mdfor checkpoint names, placement, and expected stage signals. - Read
references/troubleshooting.mdfor missing-data, missing-checkpoint, launcher, GPU-count, and version-mismatch failures. - Run
scripts/build_uniad_command.pywith--helpto render a dry-run command template without starting training or evaluation. - If you need dataset preparation or visualization instead, switch sub-skills rather than extending this one.
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.