Skip to content

README.md

Latest commit

 

History

History
60 lines (47 loc) · 2.6 KB

README.md

File metadata and controls

60 lines (47 loc) · 2.6 KB
Raw

Replication package

Drivers and figure scripts that regenerate the paper data, figures, and tables (JAX/flax + gram-form Levenberg--Marquardt).

Setup and full instructions: see the top-level README.

Files

  • <model>.py -- ensemble or repeated-solve driver; writes data to the output directory.
  • <model>_figures.py -- figures generated only from saved Parquet results.
  • collect_json.py -- writes the generated LaTeX tables and collects every number embedded in the paper into all_results.json; runs last.
  • export_paper_assets.py -- copies the figures and tables into the paper repository.
  • verify_paper_tags.py -- checks that every \jsonpath tag in the paper resolves to a leaf of all_results.json and matches the printed value.
  • utilities.py -- shared ensemble, timing, I/O, and plotting helpers.

Every driver enables JAX's persistent compilation cache (.jax_cache), so a second run of the pipeline is fully warm and its timings are compiled-cache timings. Deterministic solvers run DETERMINISTIC_TIMING_REPETITIONS times (set in utilities.py); the timing runs are saved in <model>_<experiment>_timing_df.parquet and the reported solve time is their median. NN ensembles report medians across accepted seeds: 7 seeds for the sequence-space baselines, 5 for the recursive baselines, and one for the recursive growth driver's variant experiments. The multicountry scaling ladder uses a phase-tagged timing file instead: one first solve followed by two warmed solves, with the warmed median used in its figure and both phases retained in all_results.json. The stochastic NN drivers write their policies into the output directory and the stochastic kernel drivers read them from there, so the script runs each NN driver before its kernel counterpart.

Generate everything

./generate_all_results.sh 2>&1 | tee .results/generate.log

The final step writes .results/all_results.json, the authoritative source for every number embedded in the paper. Selection and rounding to copy/paste form happen there and only there; never recompute or re-round a paper number downstream. Add a value to collect_json.py when the paper needs it. Then

uv run python -m replication.export_paper_assets
uv run python -m replication.verify_paper_tags

copy the assets into the paper repository and check the paper's tags.

Run one model

uv run python -m replication.neoclassical_growth_kernel --output_dir .results
uv run python -m replication.neoclassical_growth_kernel_figures --output_dir .results

The output directory and persistent JAX cache are git-ignored.