Skip to content

Epic/icml#2

Closed
giovannic wants to merge 3 commits into
mainfrom
epic/icml
Closed

Epic/icml#2
giovannic wants to merge 3 commits into
mainfrom
epic/icml

Conversation

@giovannic

@giovannic giovannic commented May 7, 2026

Copy link
Copy Markdown
Owner

Summary

Refocuses the the benchmark around hierarchical SBI tasks and flow/transformer-based algorithms.

New algorithms

  • TFMPE bottom-up (sbibm/algorithms/tfmpe/bottom_up.py)
  • Simformer (sbibm/algorithms/simformer/run.py)
  • DeepSet hierarchical baseline (flows, posterior, training, run).
  • FMPE wrapper added under sbibm/algorithms/sbi/.

New hierarchical tasks

Adds hierarchical_gaussian_linear (+ uniform variant), hierarchical_gaussian_mixture, hierarchical_sir, hierarchical_slcp, and hierarchical_two_moons, each with task configs and full test coverage.

Removed

  • Legacy algorithm wrappers: pyabc, pyabcranger, pyro, elfi, and most pytorch/ baselines (baseline_grid, baseline_rejection, baseline_sir, etc.).
  • sbi/mcabc, sbi/sl, sbi/smcabc.
  • Metrics dropped: ksd, mmd, moments, ppc. Added lc2st and reverse_kl.

Infrastructure

  • Hydra config tree under conf/ for experiments (ablation, benchmark, n_l_scaling), SLURM GPU launcher, and per-algorithm/task configs.
  • Move to pyproject.toml, GitHub Actions pytest workflow, Julia based ODEs removed from SIR task.
  • New README.
  • observations are now memory based instead of tracked on Git.

Tests

New task and algorithm integration tests for all hierarchical tasks. Old metric tests removed.

The hierarchical simulation based inference benchmark for the
ICML submission "Tokenised Flow Matching for hiearachical Simulation
Based Inference"
@giovannic
giovannic marked this pull request as ready for review May 18, 2026 14:34
@giovannic giovannic closed this May 20, 2026
@giovannic
giovannic deleted the epic/icml branch June 5, 2026 10:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants