60 lines
3.2 KiB
Python
60 lines
3.2 KiB
Python
"""Fixed Mold 9974ae27 objective; original regression plus unique-query E[d]."""
|
|
import numpy as np
|
|
import torch
|
|
from phase2_h2_model import weighted_loss
|
|
|
|
THETA = .03
|
|
COEFFICIENT = .001
|
|
|
|
|
|
def candidates(full_ids, query_ids, counts, coordinates, truth_seconds):
|
|
"""Exact original hard-peak reference; multiplicities do not vote in softmax.
|
|
|
|
Keep original query records separately. Repeated IDs must have identical
|
|
coordinates/truth; malformed data is an error, never an undefined waiver.
|
|
"""
|
|
full = np.asarray(full_ids)
|
|
ids = np.asarray(query_ids)
|
|
counts = np.asarray(counts)
|
|
xyz = np.asarray(coordinates, dtype=np.float64)
|
|
truth = np.asarray(truth_seconds, dtype=np.float64)
|
|
if (full.ndim != 1 or ids.ndim != 1 or not len(ids) or
|
|
full.dtype.kind not in 'iu' or ids.dtype.kind not in 'iu' or
|
|
np.any(np.diff(full) <= 0) or counts.shape != ids.shape or
|
|
counts.dtype.kind not in 'iu' or np.any(counts <= 0) or
|
|
xyz.shape != (len(ids), 3) or truth.shape != ids.shape or
|
|
not np.isfinite(xyz).all() or not np.isfinite(truth).all()):
|
|
raise ValueError('Invalid frozen query identity or truth')
|
|
unique, first, inverse = np.unique(ids, return_index=True, return_inverse=True)
|
|
if not np.array_equal(xyz, xyz[first][inverse]) or not np.array_equal(truth, truth[first][inverse]):
|
|
raise ValueError('Repeated raw ID has inconsistent coordinate/truth')
|
|
index = np.searchsorted(full, unique)
|
|
if np.any(index >= len(full)) or not np.array_equal(full[index], unique):
|
|
raise ValueError('Query ID absent from full-node geometry')
|
|
# np.unique orders raw IDs; the first true maximum implements the old tie.
|
|
t = truth[first]
|
|
peak = int(np.flatnonzero(t == t.max())[0])
|
|
distances = np.linalg.norm(xyz[first] - xyz[first][peak], axis=1)
|
|
return dict(indices=index.astype(np.int64), distances=distances,
|
|
unique_raw_ids=unique, true_peak_raw_id=int(unique[peak]),
|
|
true_peak_ties=int(np.sum(t == t.max())))
|
|
|
|
|
|
def location_loss(residual, prior64, indices, distances):
|
|
if prior64.dtype != torch.float64 or distances.dtype != torch.float64:
|
|
raise ValueError('Location prior and distances must stay FP64')
|
|
if residual.ndim != 1 or prior64.shape != residual.shape or indices.ndim != 1 or not indices.numel():
|
|
raise ValueError('Invalid location shapes')
|
|
if indices.dtype != torch.int64 or distances.shape != indices.shape or torch.unique(indices).numel() != indices.numel():
|
|
raise ValueError('Exactly one equal-weight vote per candidate')
|
|
s = prior64[indices] + residual[indices].to(torch.float64)
|
|
if not bool(torch.isfinite(s).all()) or not bool(torch.isfinite(distances).all()) or bool((distances < 0).any()):
|
|
raise ValueError('Nonfinite location data')
|
|
return torch.sum(torch.softmax((s - s.max()) / THETA, dim=0) * distances)
|
|
|
|
|
|
def objective(prediction, target, parts, nodes, total, prior64, indices, distances):
|
|
regression = weighted_loss(prediction, target, parts, nodes, total)
|
|
location = location_loss(prediction, prior64, indices, distances)
|
|
return regression + COEFFICIENT * location, regression, location
|