sampling.md

Sampling

trinity/sampling/ode.py (samplers), trinity/sampling/projection.py (state projections), trinity/solver/diffusion_placer.py (the full pipeline).

The sampler

The network predicts the clean latent x̂0 at every step; the sampler turns it into the probability-flow drift of the framing (framing.x0_to_velocity) and integrates from t = 1 (noise) to t = 1e-3:

• the time grid is linspace(1, 1e-3, nfe + 1);
• euler takes nfe explicit steps (heun is the second-order variant);
• the returned layout is the network's x̂0 read at the last time, so a run costs nfe + 1 network evaluations.

from trinity.registry import SAMPLER, build

sampler = build({"name": "euler", "num_steps": 32}, SAMPLER, framing=framing)
z = sampler.sample(model, cond, shape=(rows, max_n, 3), device="cuda")

State projections

A projection writes known answers into the sampler state. Projections are applied to the initial noise, to the state after every step, and to the returned x̂0.

key	what it writes
anchor_clamp	preplaced positions and shapes, fixed shapes, and the shape of an MIB group that has a fixed or preplaced member
mib_group_mean	one shared aspect for every all-soft MIB group: the mean of its members' ρ

Default. projections=None means DEFAULT_PROJECTIONS = ("anchor_clamp", "mib_group_mean"); this is the shipped behaviour. projections=[] is free sampling.

Paper configs. Every model in the paper is evaluated with free sampling, so every config under configs/ sets the projection list to [] explicitly. The projections are an inference-time option on a finished checkpoint, not part of the trained model.

The pipeline

DiffusionPlacer runs, per batch of cases:

1. build_group_ctx — pad len(group) × samples rows to one length and build the conditioning (features, adjacency, anchors, graph PE);
2. sample_latents — one sampler run over all rows;
3. refine_decode — the refiner (optional) and the decode to boxes;
4. legalization of every candidate (trinity.floorplan.legalize.legalize, in a process pool with legalize_workers > 0);
5. the cheapest legalized candidate per case.

raw_sample_boxes returns the first raw candidate (no refiner, no legalizer), the input of the soft metrics.