Updates benchmark documentation with latest results including README tables, speedup plots, and library metadata. Use when updating documentation, generating comparison tables, or when the user mentions update_docs.sh or documentation generation.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Updates benchmark documentation with latest results including README tables, speedup plots, and library metadata. Use when updating documentation, generating comparison tables, or when the user mentions update_docs.sh or documentation generation.
Documentation Generator
Automate updating benchmark documentation with latest results.
Quick Update
# Update all documentation
./tools/update_docs.sh
# Update with custom paths
./tools/update_docs.sh \
--image-results output/ \
--video-results output_videos/ \
--docs-dir docs/
What Gets Updated
Architecture / Policy Docs
docs/benchmark_architecture.md - Control-plane and runner architecture.
docs/benchmark_scope.md - Paper benchmark scope, transform selection, pipeline recipes, and architecture source of truth.
docs/good_plots.md - Claim-to-plot guidance for benchmark paper figures.
import pandas as pd
df = pd.read_csv('docs/images/images_speedups.csv', index_col=0)
median = df['albumentationsx'].median()
max_val = df['albumentationsx'].max()
max_transform = df['albumentationsx'].idxmax()
summary = f"AlbumentationsX is generally the fastest library for image augmentation, "
summary += f"with a median speedup of {median:.1f}× compared to other libraries. "
summary += f"For some transforms, the speedup can be as high as {max_val:.1f}× ({max_transform})."
When figure recommendations change, ensure each recommended main-text figure has a stated claim, regime, metric, support
denominator when applicable, and source CSV provenance.
Benchmark architecture docs should say that benchmark/matrix.py owns scenario/library/mode support, benchmark/policy.py
owns media defaults and slow-skip thresholds, benchmark/jobs.py owns command construction, and
benchmark/orchestrator.py owns backend dispatch.
Benchmark architecture docs should say that benchmark/config/models.py owns BenchmarkRunConfig validation,
benchmark/config/resolve.py owns YAML loading/CLI overrides/payload shaping, benchmark/config/plan.py owns dry-run
job expansion, and benchmark/config/env.py owns resolved-config metadata handoff.
Benchmark docs should use python -m benchmark.cli plan --config ... and
python -m benchmark.cli run --config ... examples for reproducible runs. Do not add flag-only benchmark run examples;
checked-in run examples live under configs/.
Benchmark docs should mention benchmark/output_naming.py for result filename policy and benchmark/cloud/paths.py for
detached GCP VM path policy whenever those rules are described.
If the benchmark matrix changes, update docs/benchmark_architecture.md, docs/benchmark_scope.md, and the relevant
skill docs in the same change.
Cloud benchmark docs should show --gcp-gcs-data-uri pointing at one dataset tarball, not a directory of individual images/videos. For macOS-created tarballs, document COPYFILE_DISABLE=1, tar --no-xattrs, and excludes for .DS_Store, AppleDouble ._*, and __MACOSX.
Micro benchmark docs should state that media is preloaded once per library and reused across transform measurements.
Pyperf docs should mention per-transform subprocess isolation, media-cache reuse, lazy transform construction, and slow-transform preflight/early-stop behavior.
Benchmark policy docs should mention lazy output materialization: micro timing should force returned outputs to contiguous memory, including contiguous NumPy conversion for Pillow/PIL Image.Image outputs. Checksums belong only in diagnostics.
Benchmark policy docs should state that library tables include only direct transform support. Missing transforms should remain unsupported instead of being recreated with benchmark-side helper code.
Environment docs should mention joined environments and cached dependency installs, including the detached GCP venv cache.
Local rerun examples should include --no-refresh-requirements when dependency versions are intentionally fixed.
