name: documentation-generator description: 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.
• .cursor/skills/benchmark-runner/SKILL.md - Agent-facing benchmark execution policy.
• .cursor/skills/paper-benchmark-execution/SKILL.md - Agent-facing paper run policy.

Image Benchmarks

• docs/images/README.md - Detailed results table
• docs/images/images_speedup_analysis.webp - Speedup visualization
• docs/images/images_speedups.csv - Raw speedup data
• README.md - Main speedup summary

Video Benchmarks

• docs/videos/README.md - Detailed results table
• docs/videos/videos_speedup_analysis.webp - Speedup visualization
• docs/videos/videos_speedups.csv - Raw speedup data
• README.md - Main speedup summary

Manual Documentation Steps

1. Generate Comparison Tables

Image benchmarks:

python tools/compare_results.py \
  --results-dir output/ \
  --update-readme docs/images/README.md

Video benchmarks:

python tools/compare_video_results.py \
  --results-dir output_videos/ \
  --update-readme docs/videos/README.md

2. Generate Speedup Plots

Image benchmarks:

python tools/generate_speedup_plots.py \
  --results-dir output/ \
  --output-dir docs/images \
  --type images \
  --reference-library albumentationsx

Video benchmarks:

python tools/generate_speedup_plots.py \
  --results-dir output_videos/ \
  --output-dir docs/videos \
  --type videos \
  --reference-library albumentationsx

3. Update Main README

The script automatically updates speedup summaries between markers:

• <!-- IMAGE_SPEEDUP_SUMMARY_START --> ... <!-- IMAGE_SPEEDUP_SUMMARY_END -->
• <!-- VIDEO_SPEEDUP_SUMMARY_START --> ... <!-- VIDEO_SPEEDUP_SUMMARY_END -->

Manual update if needed:

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})."

Library Metadata

Create metadata files for new libraries:

Image: docs/images/{library}_metadata.yaml Video: docs/videos/{library}_metadata.yaml

library_name: LibraryName
version: "1.2.3"
description: Brief description of the library
documentation: https://library.readthedocs.io
repository: https://github.com/org/library

Documentation Structure

docs/
├── images/
│   ├── README.md                      # Detailed benchmark results
│   ├── images_speedup_analysis.webp   # Main visualization
│   ├── images_speedups.csv            # Speedup data
│   ├── albumentationsx_metadata.yaml  # Library info
│   └── ...
└── videos/
    ├── README.md
    ├── videos_speedup_analysis.webp
    ├── videos_speedups.csv
    └── ...metadata.yaml files

Comparison Tools

compare_results.py (images)

python tools/compare_results.py --results-dir output/

Output format:

| Transform | albumentationsx | torchvision | kornia |
|-----------|-----------------|--------|-------------|--------|-------|
| HorizontalFlip | 1234 ± 45 | 567 ± 23 | ... | ... | ... |

compare_video_results.py (videos)

python tools/compare_video_results.py --results-dir output_videos/

Includes CPU vs GPU comparisons.

generate_speedup_plots.py

python tools/generate_speedup_plots.py \
  --results-dir output/ \
  --output-dir docs/images \
  --type images \
  --reference-library albumentationsx

Generates:

• Speedup bar chart
• CSV with speedup factors
• Statistical summary

Validation

After updating documentation:

1. Check markdown syntax:

# Tables should render correctly
# Links should be valid

1. Verify images:

ls -lh docs/images/*.webp
ls -lh docs/videos/*.webp

1. Check CSV data:

import pandas as pd
df = pd.read_csv('docs/images/images_speedups.csv', index_col=0)
print(df.head())
print(f"Shape: {df.shape}")

1. Validate README markers:

grep -n "IMAGE_SPEEDUP_SUMMARY" README.md
grep -n "VIDEO_SPEEDUP_SUMMARY" README.md

Workflow

Complete documentation update workflow:

# 1. Run benchmarks (if needed)
python -m benchmark.cli run \
  --scenario image-rgb \
  --mode micro \
  --data-dir /path/to/imagenet/val \
  --output output/rgb_micro \
  --num-items 2000

# 2. Update all documentation
./tools/update_docs.sh

# 3. Review changes
git diff README.md
git diff docs/

# 4. Commit if satisfied
git add README.md docs/
git commit -m "docs: update benchmark results"

Benchmark Policy Notes

Keep README guidance aligned with these policies:

• 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.
• 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 archive/object, not a directory of individual images.
• 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.

Troubleshooting

Missing speedup summary in README:

• Check CSV file exists: docs/images/images_speedups.csv
• Verify markers in README.md
• Run update_docs.sh again

Plot generation fails:

• Ensure matplotlib, seaborn installed: pip install -r requirements-dev.txt
• Check result files are valid JSON
• Verify all libraries have results

Table formatting issues:

• Check all result files have same transform names
• Verify no special characters in transform names
• Ensure consistent JSON structure