Skip to content

Latest commit

 

History

History
82 lines (60 loc) · 3.16 KB

baselines.md

File metadata and controls

82 lines (60 loc) · 3.16 KB

Baselines

Every baseline lives in trinity_baselines, apart from the method. The learned baselines are trained on the same data, split, augmentation, batch and step budget as Trinity, and evaluated by the same scripts (evaluation.md).

pip install -e ".[baselines]"   # torch_geometric, for the ported backbones

Direct regression

One deterministic forward from the conditioning to a layout, on the Trinity backbone (trinity_baselines.models.DirectRegressor, trained by RegressorTrainer):

head predicts
z the latent (cx/s, cy/s, ρ); areas exact by the decode
xywh the box (cx/s, cy/s, w/s, h/s); areas free
kogine run scripts/baselines/train_regressor.py    --config configs/baselines/train_reg_z.py
kogine run scripts/baselines/generate_regressor.py --config configs/baselines/generate_reg_z.py

generate_regressor.py writes the generation-cache format, so score.py, refine.py and legalize.py read it like a diffusion cache.

Ported published placers

The backbones of published learned placers, registered in BASELINE_BACKBONE and trained by the Trinity trainer when a config sets BACKBONE:

key source
flowplace_attgnn FlowPlace's AttGNN (also the ChipDiffusion backbone, trained with ddpm_cosine_eps)
diffplace_vgnn DiffPlace's VGNN
macrodiff_hetero MacroDiff+'s heterogeneous GNN
kogine run scripts/train/diffusion.py --config configs/baselines/train_flowplace.py

Their own post-processing is ported too:

  • post-hoc refiners — trinity_baselines.ports.refine.PORTS: chipd_scheduled, chipd_standard, chipd_opt (ChipDiffusion's legalizer modes), diffplace, macrodiff; selected by REFINER in scripts/eval/refine.py and legalize.py.
  • in-sampler guidance — the euler_guided sampler with GUIDANCE chipd_opt, diffplace or macrodiff:
kogine run scripts/baselines/generate_guided.py --config configs/baselines/generate_guided_chipdiffusion.py

Each ported module names its source repository and file; licenses are listed in NOTICE.

Classic solvers

solver module what it is
parsac trinity_baselines.classical.parsac Intel's PARSAC simulated annealer, driven through its Python extension
sp_sa trinity_baselines.classical.sp_sa sequence-pair simulated annealing, ported from BBOPlace-Bench

PARSAC is not vendored. Build it once:

bash scripts/baselines/build_parsac.sh   # clones IntelLabs/parsac at a pinned commit into third_party/parsac

The script applies scripts/baselines/parsac.patch (two extra knobs, both off by default) and builds the -O3 extension into third_party/parsac/build; TRINITY_PARSAC and TRINITY_PARSAC_BUILD point elsewhere.

kogine run scripts/baselines/solve_classical.py --config configs/baselines/solve_sp_sa.py
kogine run scripts/baselines/anneal_curve.py    --config configs/baselines/anneal_parsac_official.py

solve_classical.py writes a generation-style cache over the dev split; anneal_curve.py records cost against time on the official 100 cases at several budgets and seeds.