"""Audit a fixed set of original Greif geometry and immutable input arrays on CPU.""" import argparse import ast import os from pathlib import Path import numpy as np import torch import track_b_gino_greif_prepare as preparation from track_b_gino_greif_adapter import COLUMNS, GRID, deposit, gather, stencil from track_b_gino_shapenet_entry import BASE, checked, read, save, sha def fixed_cases(pilot, folds): selected = {} for fold, data in folds.items(): for role in ('learning', 'test'): cases = sorted(data['parts'][role], key=lambda case: (pilot['parts'][case]['nodes'], case)) for index in (0, len(cases) // 2, len(cases)-1): selected.setdefault(cases[index], []).append((fold, role)) return selected def geometry(pilot, case, roles): row = pilot['parts'][case] source = pilot['geometry'][case] with np.load(checked(source['path'], source['sha256']), allow_pickle=False) as archive: points, ids, raw_mass = (archive[name].copy() for name in ('points','raw_ids','mass')) def inputs(name): path = BASE / row['output'] / name return np.load(checked(path, row['output_sha256'][name]), allow_pickle=False) if len(ids) != row['nodes'] or not np.array_equal(ids, inputs('raw_ids.npy')) or not np.array_equal(raw_mass, inputs('mass.npy')): raise ValueError('Frozen original node/mass identity differs for ' + case) x = torch.from_numpy(np.asarray(inputs('features.npy'), dtype=np.float32)) mass = torch.from_numpy(raw_mass.astype(np.float32)) corner, weight = stencil(points, ids) grid, cell_mass = deposit(x, mass, corner, weight) reopened = gather(grid[:len(COLUMNS)], corner, weight) if reopened.shape != x.shape or not bool(torch.isfinite(reopened).all()): raise ValueError('Grid readout does not cover all original nodes') if (weight.sum(1) - 1.).abs().max().item() > 2e-7: raise ValueError('Incomplete trilinear node weights') ref_mass = mass.double() / mass.double().sum() source_integral = ref_mass @ x.double() scale = (ref_mass @ x.double().abs()).clamp_min(1.) grid_integral = grid[:len(COLUMNS)].reshape(len(COLUMNS),-1).double() @ cell_mass.double() relative_integral_error = ((source_integral-grid_integral).abs()/scale).max().item() relative_mass_error = abs(cell_mass.double().sum().item()-1.) if relative_mass_error > 2e-5 or relative_integral_error > 2e-5: raise ValueError('Mass-weighted transfer failed the numerical preflight') error = (reopened - x).double() column_rms = (ref_mass @ error.square()).sqrt().tolist() return dict(case=case, examined_roles=roles, original_nodes=len(ids), occupied_cells=int((cell_mass>0).sum()), mass_relative_error=relative_mass_error, field_integral_relative_error=relative_integral_error, untrained_round_trip_rms_by_column=column_rms, geometry_sha256=source['sha256'], feature_sha256=row['output_sha256']['features.npy']) def audit(): if os.environ.get('CUDA_VISIBLE_DEVICES') != '': raise PermissionError('Geometry preparation requires CUDA hidden') torch.set_num_threads(1) root = preparation.REVIEW_ROOT candidate = read(root/'preparation-binding.json') if candidate != preparation.binding() or (root/'geometry-audit.json').exists(): raise PermissionError('Only the new frozen preparation candidate') source = Path(__file__).with_name('gre24_features_v2.py') columns = next(ast.literal_eval(stmt.value) for stmt in ast.parse(source.read_text()).body if isinstance(stmt, ast.Assign) and any(isinstance(x, ast.Name) and x.id == 'COLUMNS' for x in stmt.targets)) if columns != COLUMNS: raise ValueError('Original 19-column layout differs') pilot = read(checked(preparation.PILOT, preparation.PINS[str(preparation.PILOT)])) cases = fixed_cases(pilot, candidate['folds']) rows = [geometry(pilot, case, scope) for case, scope in cases.items()] save(root/'geometry-audit.json', dict(status='PASS', candidate_sha256=sha(root/'preparation-binding.json'), size=GRID, columns=list(COLUMNS), cases=len(rows), selection='minimum, median and maximum node counts within each fold learning and test role, deduplicated', actual_heldout_labels_read=0, actual_gpu_queries=0, rows=rows)) print(len(rows), sum(r['original_nodes'] for r in rows)) if __name__ == '__main__': argparse.ArgumentParser(description=__doc__).parse_args() audit()