Skip to content

RELEASES.md

Latest commit

 

History

History
38 lines (29 loc) · 3.09 KB

RELEASES.md

File metadata and controls

38 lines (29 loc) · 3.09 KB
Raw

Ver 1.33

  • Phase-space IC sampling: training samples are no longer windows of a single origin-started orbit (which pinned every sample to the energy shell E = V0 = 2). Each potential now yields NDIFFCONFIG=4 orbits with stratified energy fraction u = (E - V(q0)) / (V0 - V(q0)) over [0.05, 0.35) / [0.35, 0.65) / [0.65, 0.95) / [0.95, 0.99). Sub-well libration and near-bottom regimes enter the training distribution (~23% of orbits cover less than 60% of the domain).
  • Explicit depth sampling: interior potential values are mapped to [V0 - D, V0 - 0.1] with D ~ U(0.5, 3.5) per sample, replacing batch-global min-max scaling (which made the depth distribution depend on batch composition).
  • Confinement by construction: E < V0 guarantees orbits stay in [0, L], removing the boundary-rejection bias of the previous pipeline. Rejection counters are printed per filter.
  • Metadata columns: parquet now carries pid, b, l, depth, E0, u per sample for OOD analysis, regime-stratified evaluation, and leakage audits. scripts/validate_data.py checks strata coverage, depth, and phase-space coverage.
  • Energy drift filter: evaluated on dense solver points relative to the per-potential energy scale (the old relative formula was unsafe for E0 < 0).
  • pytorch_template migration (v0 -> current): adds callbacks.py, checkpoint.py, provenance.py, cli.py, metrics.py; replaces config.py, main.py, pruner.py; rewrites util.py on the template base with the (V, t, q, p, ic) batch signature kept in Trainer. VaRONet variational path removed. New commands: python -m cli preflight/doctor/hpo_report. Resume via --resume and latest_model.pt.
  • configs/v1.33: run + opt configs for DeepONet/FNO/MambONet/TraONet with data/criterion/wandb fields, HPO at 10 epochs (scheduler max_iter 10), seeds [89, 231, 928].
  • Breaking: smoke mode added to data generator (neural_hamilton 3 <n>); old datasets remain loadable but lack metadata columns and are on-shell (E = 2) only.

Ver 1.31

  • Initial Condition (IC) Support: All models now require (q0, p0) as input
    • New forward signature: model(V, t, ic) where ic = (batch, 2)
    • ICEmbedding module added for conditioning
    • Additive conditioning in decoder: x = x + ic_embed(ic)
  • Autoregressive Long-term Prediction: Chain multiple windows using predicted endpoints
    • New function: autoregressive_predict(model, V, n_windows, ...)
  • Data Pipeline Updates:
    • load_data() now returns (V, t, q, p, ic) with IC extracted from q[:, 0], p[:, 0]
    • solvers.rs and true_trajectories.jl use actual IC from data (not origin)
  • Build System: just build-cargo now builds only required binaries (neural_hamilton, relevant, solvers)
  • Breaking Change: Existing checkpoints incompatible (new ic_embed layers)

Ver 1.30

  • Total data = n * NDIFFCONFIG
  • Batch size: 256

Ver 1.29

  • Remove umap

Ver 1.28

  • Introduce NDIFFCONFIG to control the number of different time configurations per potential.
    • This version uses NDIFFCONFIG=2 by default, which means two different time configurations per potential.