Skip to content

README.md

Latest commit

 

History

History
95 lines (63 loc) · 5.18 KB

README.md

File metadata and controls

95 lines (63 loc) · 5.18 KB

Scripts Folder

This folder contains developer-facing helper scripts.

Files

  • run_dev.py Starts the FastAPI app locally.

  • setup_huggingface.py Downloads and prepares the expected local Hugging Face model snapshot layout.

  • smoke_real_diffusers.py Runs a real-model smoke test through the orchestration path and writes output artifacts.

  • create_real_e2e_example.py Executes a real multi-round steering session on GPU and writes a standalone HTML walkthrough plus a trace bundle under output/examples/real_e2e_example_run/. The example is intended to demonstrate system value starting from a strong user text prompt and a clear objective.

  • run_paper_minimal_baseline_matrix.py Executes bounded paper-facing experiment bundles from a protocol YAML and writes JSON/CSV summaries plus per-run runtime bundles under the chosen paper/results/.../ directory.

  • run_paper_oracle_target_recovery.py Executes the hidden-target oracle study that starts from manual captions, steers for multiple rounds, and scores progress in CLIP space against the held-out target image.

  • run_paper_oracle_multimetric_repeated.py Repeats the oracle target-recovery study across multiple seeds per target and writes CLIP- and DINOv2-based summaries under paper/results/oracle_multimetric_repeated/.

  • run_paper_oracle_incumbent_policy_slice.py Executes the compact incumbent-policy comparison for oracle steering, contrasting carry-forward baseline, soft incumbent penalty, and hard incumbent cooldown.

  • run_paper_sampler_feedback_comparison.py Executes the controlled sampler and feedback-model comparison bundle and writes policy summaries, round curves, and paper-facing SVG plots under paper/results/sampler_feedback_comparison/.

  • run_paper_method_extension_comparison.py Executes the extended method-comparison bundle for newer samplers, richer preference models, and alternative oracle-selection policies, then writes policy summaries, round curves, and paper-facing SVG plots under paper/results/method_extension_comparison/.

  • run_paper_oracle_progress_diagnosis.py Executes a compact real-backend oracle-stagnation diagnosis study, comparing incumbent-locking baseline behavior against targeted sampler, updater, and oracle-policy fixes, then writes summary tables and a paper-facing SVG plot under paper/results/oracle_progress_diagnosis/. The same runner is also used for the newer inspired-methods slice under paper/results/oracle_inspired_methods/.

  • build_human_pairwise_eval_pack.py Builds the curated human pairwise evaluation package, including review HTML and CSV-friendly comparison artifacts under paper/results/human_pairwise_evaluation/.

  • create_configuration_samples.py Writes example YAML configuration bundles for documentation, demos, and controlled experiment setup.

  • preload_experiment_models.py Scans the paper protocol YAML files, downloads the referenced diffusion and evaluation models once into the local shared caches, and lets later experiment runs reuse them instead of refetching weights.

  • build_paper_baseline_analysis.py Reads repeated experiment CSVs from a chosen paper/results/.../tables/ directory and writes paper-facing analysis tables plus a short appendix note under the matching analysis/ directory.

  • build_journal_manuscript_assets.py Copies and prepares paper figures from repository-contained experiment bundles into paper/figures/.

  • build_journal_manuscript_html.py Renders the self-contained journal manuscript HTML with justified text and a bundled local MathJax runtime for offline equation rendering.

  • build_journal_appendix_html.py Renders the self-contained journal appendix HTML with justified text and a bundled local MathJax runtime for offline equation rendering.

  • build_paper_html.py Renders the older Markdown manuscript draft into a standalone HTML document under paper/.

  • run_e2e_debug.ps1 Launches the Playwright suite headed in Chrome for interactive debugging.

  • bootstrap.ps1 Creates a local virtual environment, installs dependencies, installs npm packages, and can optionally prepare the Hugging Face model snapshot.

  • build_release_zip.ps1 Builds an optional source release zip from tracked repository files into output/releases/.

  • build_pages_site.py Converts the repository Markdown set into a static HTML site under site/, rewrites inter-document links for GitHub Pages, and copies published image assets used by the docs.

  • generate_readme_banner.py Uses the Gemini image-generation API to create the repository banner asset used in the main README.

  • generate_doc_illustrations.py Uses the Gemini image-generation API to create conceptual illustration assets for the student tutorial and published HTML docs.

Usage

These scripts are convenience entry points for setup, release packaging, smoke testing, local development, and browser debugging.

To warm the shared model cache before running several paper experiments:

python scripts/preload_experiment_models.py --include-default-diffusion

This prepares diffusion snapshots under models/ and Hugging Face evaluation models under models/hf_cache/ so the oracle and paper runners can reuse them across runs.