"""Matched-random controls for residual temporal proposals.""" from __future__ import annotations import hashlib import json from typing import Any import numpy as np from starVLA.dataloader.infoflow_sampler import ValidAnchorManifest from starVLA.dataloader.temporal_proposal import TemporalProposal def _episode_seed(seed: int, dataset_index: int, episode_index: int) -> int: payload = f"{int(seed)}:{dataset_index}:{episode_index}".encode("utf-8") return int.from_bytes( hashlib.sha256(payload).digest()[:8], "little", signed=False, ) def build_matched_random_proposal( *, manifest: ValidAnchorManifest, residual_proposal: TemporalProposal, permutation_seed: int, ) -> TemporalProposal: """Permute residual masses within every trajectory and preserve all marginals.""" if residual_proposal.manifest_sha256 != manifest.sha256: raise ValueError("residual proposal belongs to a different manifest") if residual_proposal.score_kind != "residual": raise ValueError("matched-random control requires a residual source proposal") permuted_masses = [] for dataset_index, dataset_masses in enumerate( residual_proposal.proposal_masses ): permuted_dataset = [] for episode_index, source_mass in enumerate(dataset_masses): source = np.asarray(source_mass, dtype=np.float64) rng = np.random.default_rng( _episode_seed( permutation_seed, dataset_index, episode_index, ) ) permuted_dataset.append(source[rng.permutation(len(source))]) permuted_masses.append(permuted_dataset) identity: dict[str, Any] = { "schema_version": 1, "source_proposal_sha256": residual_proposal.sha256, "source_scorer_sha256": residual_proposal.scorer_sha256, "permutation_seed": int(permutation_seed), "permutation_scope": "within_trajectory", "mass_multiset_preserved": True, "probability_multiset_preserved": True, "importance_weight_multiset_preserved": True, } scorer_sha256 = hashlib.sha256( json.dumps( identity, sort_keys=True, separators=(",", ":"), ).encode("utf-8") ).hexdigest() proposal = TemporalProposal.from_masses( manifest=manifest, alpha=residual_proposal.alpha, masses=permuted_masses, score_kind="random", scorer_sha256=scorer_sha256, validation_certificate=identity, ) for dataset_index, (source_dataset, matched_dataset) in enumerate( zip( residual_proposal.proposal_masses, proposal.proposal_masses, ) ): for episode_index, (source_mass, matched_mass) in enumerate( zip(source_dataset, matched_dataset) ): if not np.array_equal( np.sort(np.asarray(source_mass)), np.sort(np.asarray(matched_mass)), ): raise RuntimeError( "matched-random mass multiset differs at " f"dataset={dataset_index} episode={episode_index}" ) source_q = residual_proposal.anchor_probabilities[ dataset_index ][episode_index] matched_q = proposal.anchor_probabilities[dataset_index][ episode_index ] source_w = residual_proposal.importance_weights[dataset_index][ episode_index ] matched_w = proposal.importance_weights[dataset_index][ episode_index ] if not np.allclose( np.sort(source_q), np.sort(matched_q), rtol=0.0, atol=1e-15, ): raise RuntimeError("matched-random probability multiset differs") if not np.allclose( np.sort(source_w), np.sort(matched_w), rtol=0.0, atol=1e-15, ): raise RuntimeError("matched-random weight multiset differs") return proposal