File size: 7,995 Bytes
b9301f1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 | """Hierarchical aggregation and coherent MERGE driven by toroidal phase/energy."""
from __future__ import annotations
import torch
import torch.nn as nn
import torch.nn.functional as F
class AggregationEngine(nn.Module):
def __init__(
self,
d_model: int = 256,
phase_coherence_threshold: float = 0.45,
energy_threshold: float = 0.3,
max_depth: int = 5,
merge_energy_floor: float = 0.08,
) -> None:
super().__init__()
self.d_model = d_model
self.phase_coherence_threshold = phase_coherence_threshold
self.energy_threshold = energy_threshold
self.max_depth = max_depth
self.merge_energy_floor = merge_energy_floor
def compute_coherence(self, phi: torch.Tensor, E: torch.Tensor) -> torch.Tensor:
dphi = phi[:, None, :] - phi[None, :, :]
phase = torch.cos(dphi).mean(dim=-1)
energy = torch.sqrt((E[:, None, :] * E[None, :, :]).clamp_min(0))
return phase * energy.mean(dim=-1)
def find_clusters(self, coherence: torch.Tensor, min_size: int = 2) -> list[list[int]]:
n = coherence.shape[0]
if n == 0:
return []
mask = coherence > self.phase_coherence_threshold
mask.fill_diagonal_(False)
visited = torch.zeros(n, dtype=torch.bool, device=coherence.device)
clusters: list[list[int]] = []
for i in range(n):
if visited[i]:
continue
queue = [i]
visited[i] = True
cluster: list[int] = []
while queue:
node = queue.pop()
cluster.append(node)
neighbours = torch.nonzero(mask[node] & ~visited, as_tuple=False).flatten().tolist()
for j in neighbours:
visited[j] = True
queue.append(j)
if len(cluster) >= min_size:
clusters.append(cluster)
return clusters
@staticmethod
def _circular_mean(phi: torch.Tensor, weights: torch.Tensor | None = None) -> torch.Tensor:
if weights is None:
return torch.atan2(torch.sin(phi).mean(dim=0), torch.cos(phi).mean(dim=0))
w = weights.reshape(-1, *([1] * (phi.ndim - 1)))
return torch.atan2((torch.sin(phi) * w).sum(dim=0), (torch.cos(phi) * w).sum(dim=0))
def _energy_weights(self, E: torch.Tensor, idx: torch.Tensor) -> torch.Tensor:
weights = E[idx].mean(dim=-1).clamp_min(1e-6)
return weights / weights.sum()
def _combine_cluster(
self, cluster, r, phi, omega, E, kappa, M, tau, rho
) -> dict:
"""Energy-weighted honest combine of r/φ/ω/E/κ/M/τ/ρ into one heavier atom."""
idx = torch.tensor(cluster, device=r.device, dtype=torch.long)
weights = self._energy_weights(E, idx)
r_m = F.normalize((r[idx] * weights[:, None]).sum(dim=0), dim=-1)
phi_m = self._circular_mean(phi[idx], weights)
omega_m = (omega[idx] * weights[:, None]).sum(dim=0)
e_sum = (E[idx] * weights[:, None]).sum(dim=0)
# Structural mass boost while staying in [0, 1].
e_boost = (1.0 - (1.0 - e_sum).clamp(0, 1) ** len(cluster)).clamp(0, 1)
kappa_m = (kappa[idx] * weights[:, None]).sum(dim=0)
M_m = F.normalize((M[idx] * weights[:, None]).sum(dim=0), dim=-1)
tau_m = (tau[idx] * weights[:, None]).sum(dim=0)
rho_m = rho[idx].max(dim=0).values + 1
return {
"r": r_m,
"phi": phi_m,
"omega": omega_m,
"E": e_boost,
"kappa": kappa_m,
"M": M_m,
"tau": tau_m,
"rho": rho_m,
"constituents": list(cluster),
}
def aggregate(self, r, phi, omega, E, kappa, M, tau, rho) -> tuple[list[dict], list[list[int]]]:
tensors = [x.unsqueeze(0) if x.dim() == 1 else x for x in (r, phi, omega, E, kappa, M, tau, rho)]
if tensors[0].dim() == 3 and tensors[0].shape[-2] == 1:
tensors = [x.squeeze(-2) for x in tensors]
r, phi, omega, E, kappa, M, tau, rho = tensors
if r.shape[0] < 2:
return [], []
coherence = self.compute_coherence(phi, E)
clusters = self.find_clusters(coherence)
aggregates = [
self._combine_cluster(c, r, phi, omega, E, kappa, M, tau, rho) for c in clusters
]
return aggregates, clusters
def merge_coherent(
self,
r, phi, omega, E, kappa, M, tau, rho,
*,
max_merges: int = 8,
) -> tuple[list[dict], list[int], int, dict]:
"""MERGE phase-coherent, energetic atoms into heavier structure.
Returns (merged_atom_dicts, remove_indices, merge_count, diag).
"""
tensors = [x.unsqueeze(0) if x.dim() == 1 else x for x in (r, phi, omega, E, kappa, M, tau, rho)]
if tensors[0].dim() == 3 and tensors[0].shape[-2] == 1:
tensors = [x.squeeze(-2) for x in tensors]
r, phi, omega, E, kappa, M, tau, rho = tensors
n = int(r.shape[0])
if n < 2:
return [], [], 0, {"max_phase_coherence": 0.0, "n_pairs_above_energy_floor": 0}
coherence = self.compute_coherence(phi, E)
available = torch.ones(n, dtype=torch.bool, device=r.device)
mean_E = E.mean(dim=-1)
merged: list[dict] = []
remove: list[int] = []
merge_count = 0
tri = torch.triu(coherence, diagonal=1)
# Read-only merge diagnostics: max coherence among candidate pairs
# BEFORE the phase_coherence_threshold filter (default 0.45 so mph~0.5 fires).
max_phase_coherence = float(tri.max().item()) if tri.numel() else 0.0
geo = torch.sqrt((mean_E[:, None] * mean_E[None, :]).clamp_min(0.0))
both_low = (mean_E[:, None] < self.merge_energy_floor) & (mean_E[None, :] < self.merge_energy_floor)
geo_ok = geo >= (self.merge_energy_floor * 0.5)
energy_ok = (~both_low) & geo_ok
energy_ok = torch.triu(energy_ok, diagonal=1)
n_pairs_above_energy_floor = int(energy_ok.sum().item())
flat = tri.reshape(-1)
k = min(max(n * 2, 8), flat.numel())
top_vals, top_idx = torch.topk(flat, k=k)
for score, flat_i in zip(top_vals.tolist(), top_idx.tolist()):
if merge_count >= max_merges:
break
if score < self.phase_coherence_threshold:
break
i = flat_i // n
j = flat_i % n
if i == j or not bool(available[i] and available[j]):
continue
if float(mean_E[i]) < self.merge_energy_floor and float(mean_E[j]) < self.merge_energy_floor:
continue
if float(torch.sqrt(mean_E[i] * mean_E[j])) < self.merge_energy_floor * 0.5:
continue
cluster = [i, j]
merged.append(self._combine_cluster(cluster, r, phi, omega, E, kappa, M, tau, rho))
remove.extend(cluster)
available[i] = False
available[j] = False
merge_count += 1
diag = {
"max_phase_coherence": max_phase_coherence,
"n_pairs_above_energy_floor": n_pairs_above_energy_floor,
}
return merged, sorted(set(remove)), merge_count, diag
def hierarchical_aggregate(self, r, phi, omega, E, kappa, M, tau, rho, depth: int = 3):
depth = min(depth, self.max_depth)
hierarchy = []
current = [x.clone() for x in (r, phi, omega, E, kappa, M, tau, rho)]
for _ in range(depth):
aggregates, _ = self.aggregate(*current)
if not aggregates:
break
hierarchy.append(aggregates)
if len(aggregates) < 2:
break
current = [
torch.stack([a[key] for a in aggregates])
for key in ("r", "phi", "omega", "E", "kappa", "M", "tau", "rho")
]
return hierarchy
|