Update network.py
Browse files- network.py +16 -15
network.py
CHANGED
@@ -7,13 +7,13 @@ from typing import Dict, Optional
|
|
7 |
from .node import CognitiveNode
|
8 |
|
9 |
class DynamicCognitiveNet(nn.Module):
|
10 |
-
"""
|
11 |
def __init__(self, input_size: int, output_size: int):
|
12 |
super().__init__()
|
13 |
self.input_size = input_size
|
14 |
self.output_size = output_size
|
15 |
|
16 |
-
# Node input
|
17 |
self.input_nodes = nn.ModuleList([
|
18 |
CognitiveNode(i, 1) for i in range(input_size)
|
19 |
])
|
@@ -25,7 +25,7 @@ class DynamicCognitiveNet(nn.Module):
|
|
25 |
self.connections = nn.ParameterDict()
|
26 |
self._init_base_connections()
|
27 |
|
28 |
-
#
|
29 |
self.emotional_state = nn.Parameter(torch.tensor(0.0))
|
30 |
self.optimizer = optim.AdamW(self.parameters(), lr=0.001)
|
31 |
self.loss_fn = nn.MSELoss()
|
@@ -40,6 +40,9 @@ class DynamicCognitiveNet(nn.Module):
|
|
40 |
)
|
41 |
|
42 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
|
|
|
|
43 |
# Pemrosesan input
|
44 |
activations = {}
|
45 |
for i, node in enumerate(self.input_nodes):
|
@@ -61,8 +64,8 @@ class DynamicCognitiveNet(nn.Module):
|
|
61 |
return torch.stack(outputs).squeeze()
|
62 |
|
63 |
def structural_update(self, global_reward: float):
|
64 |
-
"""Update struktur
|
65 |
-
#
|
66 |
for conn_id in list(self.connections.keys()):
|
67 |
new_weight = self.connections[conn_id] + 0.1 * global_reward
|
68 |
self.connections[conn_id].data = new_weight.clamp(-1, 1)
|
@@ -74,7 +77,7 @@ class DynamicCognitiveNet(nn.Module):
|
|
74 |
self.connections[new_conn] = nn.Parameter(torch.randn(1) * 0.1)
|
75 |
|
76 |
def _find_underutilized_connection(self) -> Optional[str]:
|
77 |
-
"""Mencari
|
78 |
input_act = {n.id: np.mean(n.recent_activations)
|
79 |
for n in self.input_nodes if n.recent_activations}
|
80 |
output_act = {n.id: np.mean(n.recent_activations)
|
@@ -92,10 +95,10 @@ class DynamicCognitiveNet(nn.Module):
|
|
92 |
self.optimizer.zero_grad()
|
93 |
|
94 |
try:
|
95 |
-
pred = self(x)
|
96 |
-
loss = self.loss_fn(pred, y)
|
97 |
-
except
|
98 |
-
print(f"Error
|
99 |
return float('nan')
|
100 |
|
101 |
# Regularisasi struktural
|
@@ -105,16 +108,14 @@ class DynamicCognitiveNet(nn.Module):
|
|
105 |
try:
|
106 |
total_loss.backward()
|
107 |
self.optimizer.step()
|
108 |
-
except
|
109 |
-
print(f"Error
|
110 |
return float('nan')
|
111 |
|
112 |
-
# Update
|
113 |
self.emotional_state.data = torch.sigmoid(
|
114 |
self.emotional_state + (0.5 - loss.item()) * 0.1
|
115 |
)
|
116 |
-
|
117 |
-
# Update struktur
|
118 |
self.structural_update(0.5 - loss.item())
|
119 |
|
120 |
return total_loss.item()
|
|
|
7 |
from .node import CognitiveNode
|
8 |
|
9 |
class DynamicCognitiveNet(nn.Module):
|
10 |
+
"""Arsitektur jaringan dengan manajemen tensor yang robust"""
|
11 |
def __init__(self, input_size: int, output_size: int):
|
12 |
super().__init__()
|
13 |
self.input_size = input_size
|
14 |
self.output_size = output_size
|
15 |
|
16 |
+
# Node dengan input size 1
|
17 |
self.input_nodes = nn.ModuleList([
|
18 |
CognitiveNode(i, 1) for i in range(input_size)
|
19 |
])
|
|
|
25 |
self.connections = nn.ParameterDict()
|
26 |
self._init_base_connections()
|
27 |
|
28 |
+
# Sistem pembelajaran
|
29 |
self.emotional_state = nn.Parameter(torch.tensor(0.0))
|
30 |
self.optimizer = optim.AdamW(self.parameters(), lr=0.001)
|
31 |
self.loss_fn = nn.MSELoss()
|
|
|
40 |
)
|
41 |
|
42 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
43 |
+
# Validasi dimensi input
|
44 |
+
x = x.view(-1)
|
45 |
+
|
46 |
# Pemrosesan input
|
47 |
activations = {}
|
48 |
for i, node in enumerate(self.input_nodes):
|
|
|
64 |
return torch.stack(outputs).squeeze()
|
65 |
|
66 |
def structural_update(self, global_reward: float):
|
67 |
+
"""Update struktur jaringan"""
|
68 |
+
# Update kekuatan koneksi
|
69 |
for conn_id in list(self.connections.keys()):
|
70 |
new_weight = self.connections[conn_id] + 0.1 * global_reward
|
71 |
self.connections[conn_id].data = new_weight.clamp(-1, 1)
|
|
|
77 |
self.connections[new_conn] = nn.Parameter(torch.randn(1) * 0.1)
|
78 |
|
79 |
def _find_underutilized_connection(self) -> Optional[str]:
|
80 |
+
"""Mencari pasangan node yang kurang aktif"""
|
81 |
input_act = {n.id: np.mean(n.recent_activations)
|
82 |
for n in self.input_nodes if n.recent_activations}
|
83 |
output_act = {n.id: np.mean(n.recent_activations)
|
|
|
95 |
self.optimizer.zero_grad()
|
96 |
|
97 |
try:
|
98 |
+
pred = self(x.view(-1))
|
99 |
+
loss = self.loss_fn(pred, y.view(-1))
|
100 |
+
except Exception as e:
|
101 |
+
print(f"Error forward: {e}")
|
102 |
return float('nan')
|
103 |
|
104 |
# Regularisasi struktural
|
|
|
108 |
try:
|
109 |
total_loss.backward()
|
110 |
self.optimizer.step()
|
111 |
+
except Exception as e:
|
112 |
+
print(f"Error backward: {e}")
|
113 |
return float('nan')
|
114 |
|
115 |
+
# Update emosi
|
116 |
self.emotional_state.data = torch.sigmoid(
|
117 |
self.emotional_state + (0.5 - loss.item()) * 0.1
|
118 |
)
|
|
|
|
|
119 |
self.structural_update(0.5 - loss.item())
|
120 |
|
121 |
return total_loss.item()
|