| | import torch |
| | import torch.nn as nn |
| | import torch.nn.functional as F |
| | from collections import deque |
| | from typing import Deque, Dict, Any |
| |
|
| | class CognitiveMemory(nn.Module): |
| | """Memory system dengan manajemen dimensi yang ketat""" |
| | def __init__(self, context_size: int, capacity: int = 100): |
| | super().__init__() |
| | self.context_size = context_size |
| | self.capacity = capacity |
| | self.memory_queue: Deque[Dict[str, Any]] = deque(maxlen=capacity) |
| | |
| | |
| | self.key_proj = nn.Linear(context_size, context_size) |
| | self.value_proj = nn.Linear(context_size, context_size) |
| | self.importance_decay = nn.Parameter(torch.tensor(0.95)) |
| |
|
| | def add_memory(self, context: torch.Tensor, activation: float): |
| | """Menyimpan memori dengan dimensi terkontrol""" |
| | importance = torch.sigmoid(torch.tensor(activation * 0.5 + 0.2)) |
| | self.memory_queue.append({ |
| | 'context': context.detach().clone().squeeze(), |
| | 'importance': importance, |
| | 'age': torch.tensor(0.0) |
| | }) |
| |
|
| | def consolidate_memories(self): |
| | """Konsolidasi memori dengan validasi dimensi""" |
| | self.memory_queue = deque( |
| | [m for m in self.memory_queue if m['importance'] > 0.2], |
| | maxlen=self.capacity |
| | ) |
| |
|
| | def retrieve(self, query: torch.Tensor) -> torch.Tensor: |
| | """Retrieval dengan penanganan tensor 1D""" |
| | if not self.memory_queue: |
| | return torch.zeros(self.context_size, device=query.device) |
| | |
| | |
| | contexts = torch.stack([m['context'] for m in self.memory_queue]) |
| | keys = self.key_proj(contexts) |
| | values = self.value_proj(contexts) |
| | query_proj = self.key_proj(query.squeeze()) |
| | |
| | scores = F.softmax(keys @ query_proj, dim=0) |
| | return (scores.unsqueeze(1) * values).sum(dim=0) |
