Skip to content

Figure 17 — Run profiles

fig17_profiles

Vector PDF

File: fig17_profiles.pdf

Suggested caption. The three run profiles. min (default, intended for Colab) shares one trained generator per dataset across seeds. mid retrains the VAE on every seed and adds the noise twin vae_rand_b. max uses the full series, retrains generators on every seed, and turns on hyperparameter search and the decision-transfer grid.

Notes for the text

  • What a seed varies depends on the profile. In max, VAE and Diffusion-TS are retrained on every seed, so per-dataset conclusions generalise over the training procedure. In min, seeds vary only pool sampling, random paths and model fitting, so conclusions are conditional on the trained generator. mid sits in between: the VAE is retrained on every seed, Diffusion-TS once per dataset.
  • min compares 16 methods, mid 17 (+ the noise twin vae_rand_b), max 35: in max every seam processing also runs over every assembler except the random path and vae_quilting, not only vae_zseries (smooth_paths="all"). Each count includes the 4 published results of frozen_methods (vae_zseries and its seam-processing variants), which the pipeline does not run. Without vae_rand_b (min) the noise is the between-seed spread of vae_rand. All profiles use the same 9 datasets and 10 seeds.
  • Profiles write to separate data and artifact folders, so their results never mix. Task thresholds come from target_params.json (min, mid) or target_params_max.json (max), because the min train split falls inside val/OOT of max.
  • Source: marathonts/bench/config.py (Config.profile), BENCHMARK.md.