Upload mle/energy.py
Browse files- mle/energy.py +347 -0
mle/energy.py
ADDED
|
@@ -0,0 +1,347 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Paysage d'Énergie Apprenant
|
| 3 |
+
|
| 4 |
+
La fonction d'énergie évalue la cohérence d'un état du système.
|
| 5 |
+
Elle doit être :
|
| 6 |
+
- Locale : les mises à jour ne dépendent que des voisins
|
| 7 |
+
- Apprenante : s'ajuste avec l'expérience
|
| 8 |
+
- Discriminative : basse énergie = cohérent, haute = incohérent
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
import numpy as np
|
| 12 |
+
from numba import njit, prange
|
| 13 |
+
from typing import Dict, List, Tuple, Optional, Set
|
| 14 |
+
import logging
|
| 15 |
+
|
| 16 |
+
logger = logging.getLogger(__name__)
|
| 17 |
+
|
| 18 |
+
VECTOR_SIZE = 4096
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def compute_local_hamming_energy_fast(
|
| 22 |
+
state: np.ndarray,
|
| 23 |
+
neighbors: np.ndarray,
|
| 24 |
+
weights: np.ndarray,
|
| 25 |
+
) -> float:
|
| 26 |
+
"""
|
| 27 |
+
Énergie locale de Hamming : somme pondérée des distances aux voisins.
|
| 28 |
+
Version numpy vectorisée rapide.
|
| 29 |
+
"""
|
| 30 |
+
N = neighbors.shape[0]
|
| 31 |
+
if N == 0:
|
| 32 |
+
return 0.0
|
| 33 |
+
# XOR vectorisé : state ^ neighbors[i]
|
| 34 |
+
xor_result = state.astype(np.uint8) ^ neighbors # (N, 4096)
|
| 35 |
+
# Somme par ligne
|
| 36 |
+
dists = np.sum(xor_result, axis=1)
|
| 37 |
+
return float(np.dot(weights, dists)) / N
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
def compute_bit_flip_delta_fast(
|
| 41 |
+
state: np.ndarray,
|
| 42 |
+
neighbors: np.ndarray,
|
| 43 |
+
weights: np.ndarray,
|
| 44 |
+
biases: np.ndarray,
|
| 45 |
+
) -> np.ndarray:
|
| 46 |
+
"""
|
| 47 |
+
Calcule le delta d'énergie pour chaque flip de bit possible.
|
| 48 |
+
Version numpy vectorisée.
|
| 49 |
+
"""
|
| 50 |
+
N = neighbors.shape[0]
|
| 51 |
+
if N == 0:
|
| 52 |
+
return biases.copy()
|
| 53 |
+
|
| 54 |
+
# Pour chaque bit, calculer le changement si on le flippe
|
| 55 |
+
# dist_actuel = sum(weights[i] * (state[j] ^ neighbors[i,j]))
|
| 56 |
+
# dist_nouveau = sum(weights[i] * ((1-state[j]) ^ neighbors[i,j]))
|
| 57 |
+
|
| 58 |
+
state_expanded = state[np.newaxis, :] # (1, 4096)
|
| 59 |
+
# XOR actuel
|
| 60 |
+
xor_current = state_expanded ^ neighbors # (N, 4096)
|
| 61 |
+
# XOR après flip
|
| 62 |
+
flipped = 1 - state_expanded
|
| 63 |
+
xor_flipped = flipped ^ neighbors # (N, 4096)
|
| 64 |
+
|
| 65 |
+
# Delta = new_dist - old_dist pour chaque bit
|
| 66 |
+
delta_xor = xor_flipped.astype(np.float64) - xor_current.astype(np.float64) # (N, 4096)
|
| 67 |
+
# Pondération
|
| 68 |
+
weights_col = weights[:, np.newaxis] # (N, 1)
|
| 69 |
+
weighted_delta = delta_xor * weights_col # (N, 4096)
|
| 70 |
+
|
| 71 |
+
deltas = np.sum(weighted_delta, axis=0) # (4096,)
|
| 72 |
+
deltas += biases * (2.0 * (1 - state.astype(np.float64)) - 1.0)
|
| 73 |
+
|
| 74 |
+
return deltas
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
class EnergyLandscape:
|
| 78 |
+
"""
|
| 79 |
+
Paysage d'énergie apprenant avec mises à jour locales.
|
| 80 |
+
|
| 81 |
+
Components:
|
| 82 |
+
- hamming_weight: poids de la distance de Hamming
|
| 83 |
+
- association_weight: poids des associations Hebbian
|
| 84 |
+
- structure_weight: poids de la cohérence structurelle (binding)
|
| 85 |
+
- biases: biais par bit (apprenant)
|
| 86 |
+
- local_weights: poids par voisin (apprenant)
|
| 87 |
+
|
| 88 |
+
Apprentissage:
|
| 89 |
+
- Renforcement des associations dans les états de basse énergie
|
| 90 |
+
- Affaiblissement dans les états instables
|
| 91 |
+
- Ajustement des biais locaux
|
| 92 |
+
"""
|
| 93 |
+
|
| 94 |
+
def __init__(
|
| 95 |
+
self,
|
| 96 |
+
hamming_weight: float = 1.0,
|
| 97 |
+
association_weight: float = 0.5,
|
| 98 |
+
structure_weight: float = 0.3,
|
| 99 |
+
bias_learning_rate: float = 0.01,
|
| 100 |
+
association_decay: float = 0.99,
|
| 101 |
+
energy_threshold_low: float = 500.0,
|
| 102 |
+
energy_threshold_high: float = 1500.0,
|
| 103 |
+
):
|
| 104 |
+
self.hamming_weight = hamming_weight
|
| 105 |
+
self.association_weight = association_weight
|
| 106 |
+
self.structure_weight = structure_weight
|
| 107 |
+
self.bias_lr = bias_learning_rate
|
| 108 |
+
self.assoc_decay = association_decay
|
| 109 |
+
self.low_threshold = energy_threshold_low
|
| 110 |
+
self.high_threshold = energy_threshold_high
|
| 111 |
+
|
| 112 |
+
# Biases par bit (apprenant)
|
| 113 |
+
self.biases = np.zeros(VECTOR_SIZE, dtype=np.float64)
|
| 114 |
+
|
| 115 |
+
# Associations apprises : (id1, id2) -> strength
|
| 116 |
+
self.associations: Dict[Tuple[int, int], float] = {}
|
| 117 |
+
|
| 118 |
+
# Poids structurels pour binding
|
| 119 |
+
self.structure_weights: Dict[Tuple[int, int], float] = {}
|
| 120 |
+
|
| 121 |
+
# Historique pour apprentissage
|
| 122 |
+
self.energy_history: List[float] = []
|
| 123 |
+
self.min_energy_history: List[float] = []
|
| 124 |
+
self.stable_states: List[np.ndarray] = []
|
| 125 |
+
|
| 126 |
+
# Stats
|
| 127 |
+
self.n_updates = 0
|
| 128 |
+
self.total_energy = 0.0
|
| 129 |
+
|
| 130 |
+
def compute_energy(
|
| 131 |
+
self,
|
| 132 |
+
state: np.ndarray,
|
| 133 |
+
neighbor_vectors: np.ndarray,
|
| 134 |
+
neighbor_ids: List[int],
|
| 135 |
+
neighbor_metadata: Optional[List] = None,
|
| 136 |
+
) -> float:
|
| 137 |
+
"""
|
| 138 |
+
Calcule l'énergie totale d'un état.
|
| 139 |
+
|
| 140 |
+
Args:
|
| 141 |
+
state: vecteur courant (4096,) uint8
|
| 142 |
+
neighbor_vectors: (N, 4096) voisins actifs
|
| 143 |
+
neighbor_ids: IDs des voisins
|
| 144 |
+
neighbor_metadata: métadonnées optionnelles
|
| 145 |
+
|
| 146 |
+
Returns:
|
| 147 |
+
énergie scalaire (plus bas = plus cohérent)
|
| 148 |
+
"""
|
| 149 |
+
N = len(neighbor_ids) if neighbor_ids else 0
|
| 150 |
+
if N == 0:
|
| 151 |
+
return float(np.sum(self.biases * (2.0 * state - 1.0)))
|
| 152 |
+
|
| 153 |
+
# Poids des voisins (apprenant)
|
| 154 |
+
weights = np.array([
|
| 155 |
+
self.associations.get(
|
| 156 |
+
tuple(sorted((-1, nid))), 0.5 # Default weight
|
| 157 |
+
) + 0.5
|
| 158 |
+
for nid in neighbor_ids
|
| 159 |
+
], dtype=np.float64)
|
| 160 |
+
|
| 161 |
+
# 1. Énergie de Hamming (vectorisée rapide)
|
| 162 |
+
hamming_energy = compute_local_hamming_energy_fast(
|
| 163 |
+
state, neighbor_vectors, weights
|
| 164 |
+
)
|
| 165 |
+
hamming_energy *= self.hamming_weight
|
| 166 |
+
|
| 167 |
+
# 2. Énergie d'association (Hebbian-like)
|
| 168 |
+
association_energy = 0.0
|
| 169 |
+
if len(neighbor_ids) > 1:
|
| 170 |
+
for i in range(len(neighbor_ids)):
|
| 171 |
+
for j in range(i+1, len(neighbor_ids)):
|
| 172 |
+
pair = tuple(sorted((neighbor_ids[i], neighbor_ids[j])))
|
| 173 |
+
strength = self.associations.get(pair, 0.0)
|
| 174 |
+
# Distance entre voisins : proches = faible énergie
|
| 175 |
+
dist = np.sum(neighbor_vectors[i] != neighbor_vectors[j])
|
| 176 |
+
association_energy += strength * dist
|
| 177 |
+
|
| 178 |
+
association_energy *= self.association_weight
|
| 179 |
+
|
| 180 |
+
# 3. Énergie de structure (binding)
|
| 181 |
+
structure_energy = 0.0
|
| 182 |
+
if len(neighbor_ids) > 1:
|
| 183 |
+
for i in range(len(neighbor_ids)):
|
| 184 |
+
for j in range(i+1, len(neighbor_ids)):
|
| 185 |
+
pair = tuple(sorted((neighbor_ids[i], neighbor_ids[j])))
|
| 186 |
+
struct_weight = self.structure_weights.get(pair, 0.0)
|
| 187 |
+
# Corrélation comme proxy de structure
|
| 188 |
+
corr = np.mean(neighbor_vectors[i] == neighbor_vectors[j])
|
| 189 |
+
structure_energy += struct_weight * (1.0 - corr)
|
| 190 |
+
|
| 191 |
+
structure_energy *= self.structure_weight
|
| 192 |
+
|
| 193 |
+
# 4. Biais local
|
| 194 |
+
bias_energy = float(np.sum(self.biases * (2.0 * state.astype(np.float64) - 1.0)))
|
| 195 |
+
|
| 196 |
+
total = hamming_energy + association_energy + structure_energy + bias_energy
|
| 197 |
+
|
| 198 |
+
self.energy_history.append(total)
|
| 199 |
+
if len(self.energy_history) > 1000:
|
| 200 |
+
self.energy_history = self.energy_history[-1000:]
|
| 201 |
+
|
| 202 |
+
return total
|
| 203 |
+
|
| 204 |
+
def get_bit_flip_deltas(
|
| 205 |
+
self,
|
| 206 |
+
state: np.ndarray,
|
| 207 |
+
neighbor_vectors: np.ndarray,
|
| 208 |
+
neighbor_ids: List[int],
|
| 209 |
+
) -> np.ndarray:
|
| 210 |
+
"""
|
| 211 |
+
Calcule le delta d'énergie pour chaque bit flip possible.
|
| 212 |
+
Utilisé par l'inférence pour trouver les meilleurs flips.
|
| 213 |
+
|
| 214 |
+
Returns:
|
| 215 |
+
deltas: (4096,) négatif = flip réduit l'énergie
|
| 216 |
+
"""
|
| 217 |
+
N = len(neighbor_ids)
|
| 218 |
+
if N == 0:
|
| 219 |
+
# Juste les biais
|
| 220 |
+
return self.biases.copy()
|
| 221 |
+
|
| 222 |
+
weights = np.array([
|
| 223 |
+
self.associations.get(tuple(sorted((-1, nid))), 0.5) + 0.5
|
| 224 |
+
for nid in neighbor_ids
|
| 225 |
+
], dtype=np.float64)
|
| 226 |
+
|
| 227 |
+
return compute_bit_flip_delta_fast(state, neighbor_vectors, weights, self.biases)
|
| 228 |
+
|
| 229 |
+
def update_from_state(
|
| 230 |
+
self,
|
| 231 |
+
state: np.ndarray,
|
| 232 |
+
neighbor_ids: List[int],
|
| 233 |
+
energy: float,
|
| 234 |
+
is_stable: bool = False,
|
| 235 |
+
):
|
| 236 |
+
"""
|
| 237 |
+
Met à jour le paysage d'énergie à partir d'un état observé.
|
| 238 |
+
Appelé après chaque itération d'inférence.
|
| 239 |
+
|
| 240 |
+
Règles:
|
| 241 |
+
- Basse énergie stable : renforce les associations
|
| 242 |
+
- Haute énergie instable : affaiblit ou réorganise
|
| 243 |
+
"""
|
| 244 |
+
self.n_updates += 1
|
| 245 |
+
self.total_energy += energy
|
| 246 |
+
|
| 247 |
+
if len(neighbor_ids) < 2:
|
| 248 |
+
return
|
| 249 |
+
|
| 250 |
+
# Normalise l'énergie relative
|
| 251 |
+
if len(self.energy_history) > 20:
|
| 252 |
+
recent_mean = np.mean(self.energy_history[-20:])
|
| 253 |
+
recent_std = np.std(self.energy_history[-20:]) + 1e-8
|
| 254 |
+
normalized_energy = (energy - recent_mean) / recent_std
|
| 255 |
+
else:
|
| 256 |
+
normalized_energy = 0.0
|
| 257 |
+
|
| 258 |
+
# Détermine si l'état est cohérent ou incohérent
|
| 259 |
+
is_coherent = normalized_energy < -0.5 or (energy < self.low_threshold and is_stable)
|
| 260 |
+
is_incoherent = normalized_energy > 0.5 or energy > self.high_threshold
|
| 261 |
+
|
| 262 |
+
# Met à jour les associations
|
| 263 |
+
for i in range(len(neighbor_ids)):
|
| 264 |
+
for j in range(i+1, len(neighbor_ids)):
|
| 265 |
+
pair = tuple(sorted((neighbor_ids[i], neighbor_ids[j])))
|
| 266 |
+
|
| 267 |
+
if is_coherent:
|
| 268 |
+
# Renforce l'association (Hebbian-like)
|
| 269 |
+
current = self.associations.get(pair, 0.0)
|
| 270 |
+
self.associations[pair] = min(1.0, current + self.bias_lr * (1.0 - current))
|
| 271 |
+
elif is_incoherent:
|
| 272 |
+
# Affaiblit l'association (anti-Hebbian)
|
| 273 |
+
current = self.associations.get(pair, 0.0)
|
| 274 |
+
self.associations[pair] = max(-0.5, current - self.bias_lr * 2.0)
|
| 275 |
+
|
| 276 |
+
# Met à jour les biais locaux
|
| 277 |
+
state_float = state.astype(np.float64)
|
| 278 |
+
if is_coherent:
|
| 279 |
+
# Renforce les bits qui sont cohérents avec le voisinage
|
| 280 |
+
self.biases += self.bias_lr * (2.0 * state_float - 1.0) * 0.1
|
| 281 |
+
elif is_incoherent:
|
| 282 |
+
# Affaiblit les bits qui sont incohérents
|
| 283 |
+
self.biases -= self.bias_lr * (2.0 * state_float - 1.0) * 0.2
|
| 284 |
+
|
| 285 |
+
# Normalise les biais pour éviter la dérive
|
| 286 |
+
self.biases = np.clip(self.biases, -2.0, 2.0)
|
| 287 |
+
|
| 288 |
+
# Si stable et basse énergie, mémorise l'état
|
| 289 |
+
if is_stable and is_coherent:
|
| 290 |
+
self.stable_states.append(state.copy())
|
| 291 |
+
if len(self.stable_states) > 100:
|
| 292 |
+
self.stable_states = self.stable_states[-100:]
|
| 293 |
+
|
| 294 |
+
# Déduit les associations avec le temps
|
| 295 |
+
if self.n_updates % 100 == 0:
|
| 296 |
+
self._decay_associations()
|
| 297 |
+
|
| 298 |
+
def update_structure_weights(
|
| 299 |
+
self,
|
| 300 |
+
bound_state: np.ndarray,
|
| 301 |
+
component_ids: List[int],
|
| 302 |
+
coherence: float,
|
| 303 |
+
):
|
| 304 |
+
"""
|
| 305 |
+
Met à jour les poids structurels pour le binding.
|
| 306 |
+
"""
|
| 307 |
+
for i in range(len(component_ids)):
|
| 308 |
+
for j in range(i+1, len(component_ids)):
|
| 309 |
+
pair = tuple(sorted((component_ids[i], component_ids[j])))
|
| 310 |
+
current = self.structure_weights.get(pair, 0.0)
|
| 311 |
+
# Mise à jour basée sur la cohérence
|
| 312 |
+
if coherence > 0.7:
|
| 313 |
+
self.structure_weights[pair] = min(1.0, current + self.bias_lr)
|
| 314 |
+
elif coherence < 0.3:
|
| 315 |
+
self.structure_weights[pair] = max(-0.5, current - self.bias_lr * 1.5)
|
| 316 |
+
|
| 317 |
+
def _decay_associations(self):
|
| 318 |
+
"""Décroissance périodique des associations peu utilisées."""
|
| 319 |
+
to_remove = []
|
| 320 |
+
for pair, strength in self.associations.items():
|
| 321 |
+
decayed = strength * self.assoc_decay
|
| 322 |
+
if abs(decayed) < 0.01:
|
| 323 |
+
to_remove.append(pair)
|
| 324 |
+
else:
|
| 325 |
+
self.associations[pair] = decayed
|
| 326 |
+
|
| 327 |
+
for pair in to_remove:
|
| 328 |
+
del self.associations[pair]
|
| 329 |
+
|
| 330 |
+
def get_association_strength(self, id1: int, id2: int) -> float:
|
| 331 |
+
"""Retourne la force d'association entre deux vecteurs."""
|
| 332 |
+
pair = tuple(sorted((id1, id2)))
|
| 333 |
+
return self.associations.get(pair, 0.0)
|
| 334 |
+
|
| 335 |
+
def get_stats(self) -> Dict:
|
| 336 |
+
n_assoc = len(self.associations)
|
| 337 |
+
n_struct = len(self.structure_weights)
|
| 338 |
+
recent_energy = np.mean(self.energy_history[-100:]) if self.energy_history else 0.0
|
| 339 |
+
|
| 340 |
+
return {
|
| 341 |
+
'n_associations': n_assoc,
|
| 342 |
+
'n_structure_weights': n_struct,
|
| 343 |
+
'mean_bias': float(np.mean(np.abs(self.biases))),
|
| 344 |
+
'recent_mean_energy': float(recent_energy),
|
| 345 |
+
'n_updates': self.n_updates,
|
| 346 |
+
'n_stable_states': len(self.stable_states),
|
| 347 |
+
}
|