Update core/graph_mamba.py
Browse files- core/graph_mamba.py +299 -219
core/graph_mamba.py
CHANGED
@@ -8,206 +8,173 @@ import logging
|
|
8 |
|
9 |
logger = logging.getLogger(__name__)
|
10 |
|
11 |
-
class
|
12 |
-
"""
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
super().__init__()
|
15 |
self.d_model = d_model
|
|
|
|
|
16 |
|
17 |
-
#
|
18 |
-
self.
|
19 |
-
self.
|
20 |
-
self.
|
21 |
-
self.register_buffer('potential_energy', torch.zeros(d_model))
|
22 |
-
|
23 |
-
# Field interactions
|
24 |
-
self.attraction_projection = nn.Linear(d_model, d_model)
|
25 |
-
self.repulsion_projection = nn.Linear(d_model, d_model)
|
26 |
|
27 |
-
#
|
28 |
-
self.
|
29 |
-
self.
|
|
|
30 |
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
|
|
35 |
|
36 |
-
|
37 |
-
current_velocity = self.momentum_vectors / (self.cognitive_mass + 1e-8)
|
38 |
-
new_velocity = current_velocity + acceleration * dt
|
39 |
|
40 |
-
|
41 |
-
|
42 |
|
43 |
-
#
|
44 |
-
|
|
|
45 |
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
|
52 |
-
#
|
53 |
-
|
54 |
|
55 |
-
#
|
56 |
-
|
|
|
57 |
|
58 |
-
return
|
59 |
|
60 |
-
def
|
61 |
-
"""
|
62 |
-
|
63 |
-
x = x.unsqueeze(0)
|
64 |
-
batch_size, seq_len, d_model = x.shape
|
65 |
|
66 |
-
#
|
67 |
-
|
68 |
-
|
|
|
69 |
|
70 |
-
#
|
71 |
-
|
|
|
|
|
72 |
|
73 |
-
#
|
74 |
-
|
|
|
75 |
|
76 |
-
#
|
77 |
-
|
78 |
-
|
|
|
|
|
79 |
|
80 |
-
return
|
81 |
|
82 |
-
class
|
83 |
-
"""
|
84 |
-
def __init__(self, d_model
|
85 |
super().__init__()
|
86 |
self.d_model = d_model
|
87 |
-
self.d_astrocyte = int(d_model * astrocyte_ratio)
|
88 |
-
|
89 |
-
# Fast neuronal processing
|
90 |
-
self.neuron_fast = nn.Linear(d_model, d_model)
|
91 |
-
self.neuron_dropout = nn.Dropout(0.1)
|
92 |
-
|
93 |
-
# Slow astrocyte processing
|
94 |
-
self.astrocyte_slow = nn.Linear(d_model, self.d_astrocyte)
|
95 |
-
self.astrocyte_integration = nn.Linear(self.d_astrocyte, d_model)
|
96 |
-
self.astrocyte_dropout = nn.Dropout(0.1)
|
97 |
|
98 |
-
#
|
99 |
-
self.
|
|
|
100 |
|
101 |
-
#
|
102 |
-
self.
|
103 |
-
self.
|
104 |
-
|
105 |
-
# Memory for slow dynamics
|
106 |
-
self.register_buffer('astrocyte_memory', torch.zeros(1, self.d_astrocyte))
|
107 |
-
self.memory_decay = 0.9
|
108 |
|
109 |
def forward(self, x):
|
110 |
-
batch_size = x.size(0) if x.dim() == 3 else 1
|
111 |
if x.dim() == 2:
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
self.
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
# Slow astrocyte integration
|
124 |
-
astrocyte_input = self.astrocyte_slow(x_momentum)
|
125 |
-
self.astrocyte_memory = self.memory_decay * self.astrocyte_memory + (1 - self.memory_decay) * astrocyte_input.mean(dim=1)
|
126 |
-
slow_out = self.astrocyte_dropout(torch.tanh(self.astrocyte_integration(self.astrocyte_memory))).unsqueeze(1).expand(-1, x.size(1), -1)
|
127 |
-
|
128 |
-
# Multi-timescale gating
|
129 |
-
fast_gate = torch.sigmoid(self.fast_gate(x_momentum))
|
130 |
-
slow_gate = torch.sigmoid(self.slow_gate(self.astrocyte_memory)).unsqueeze(1).expand(-1, x.size(1), -1)
|
131 |
-
|
132 |
-
# Combine with momentum
|
133 |
-
output = fast_gate * fast_out + slow_gate * slow_out
|
134 |
-
|
135 |
-
return output.squeeze(0) if output.size(0) == 1 else output
|
136 |
|
137 |
-
class
|
138 |
-
"""
|
139 |
-
def __init__(self, d_model
|
140 |
super().__init__()
|
141 |
self.d_model = d_model
|
142 |
-
self.
|
143 |
-
self.d_state = d_state
|
144 |
|
145 |
-
|
146 |
-
self.
|
147 |
-
self.
|
148 |
-
self.dt_proj = nn.Linear(1, self.d_inner, bias=True)
|
149 |
|
150 |
-
#
|
151 |
-
|
152 |
-
self.
|
153 |
-
self.
|
154 |
-
self.out_proj = nn.Linear(self.d_inner, d_model, bias=False)
|
155 |
-
|
156 |
-
# Energy conservation
|
157 |
-
self.energy_projection = nn.Linear(d_model, d_model)
|
158 |
|
159 |
-
|
160 |
-
|
161 |
-
x = x.unsqueeze(0)
|
162 |
|
163 |
-
|
|
|
|
|
164 |
|
165 |
-
|
166 |
-
|
|
|
167 |
|
168 |
-
|
169 |
-
x_inner, z = xz.chunk(2, dim=-1)
|
170 |
|
171 |
-
#
|
172 |
-
|
173 |
-
x_inner = self.conv1d(x_inner)[:, :, :length]
|
174 |
-
x_inner = x_inner.transpose(1, 2)
|
175 |
-
x_inner = F.silu(x_inner)
|
176 |
-
|
177 |
-
# State space with physics
|
178 |
-
y = self.selective_scan(x_inner)
|
179 |
-
y = y * F.silu(z)
|
180 |
|
181 |
-
#
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
return output
|
188 |
-
|
189 |
-
def selective_scan(self, x):
|
190 |
-
batch, length, d_inner = x.shape
|
191 |
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
|
196 |
-
|
197 |
-
deltaB = delta.unsqueeze(-1) * B.unsqueeze(2)
|
198 |
-
|
199 |
-
states = torch.zeros(batch, d_inner, self.d_state, device=x.device)
|
200 |
-
outputs = []
|
201 |
-
|
202 |
-
for i in range(length):
|
203 |
-
states = deltaA[:, i] * states + deltaB[:, i] * x[:, i, :, None]
|
204 |
-
y = (states @ C[:, i, :, None]).squeeze(-1) + self.D * x[:, i]
|
205 |
-
outputs.append(y)
|
206 |
-
|
207 |
-
return torch.stack(outputs, dim=1)
|
208 |
|
209 |
-
class
|
210 |
-
"""
|
211 |
def __init__(self, config):
|
212 |
super().__init__()
|
213 |
|
@@ -219,129 +186,243 @@ class CognitiveMambaGraphMamba(nn.Module):
|
|
219 |
# Input processing
|
220 |
self.input_proj = nn.Linear(input_dim, d_model)
|
221 |
self.input_norm = nn.LayerNorm(d_model)
|
|
|
222 |
|
223 |
-
#
|
|
|
|
|
|
|
224 |
self.gcn_layers = nn.ModuleList([
|
225 |
GCNConv(d_model, d_model) for _ in range(n_layers)
|
226 |
])
|
227 |
|
228 |
-
# Revolutionary components
|
229 |
self.astrocyte_layers = nn.ModuleList([
|
230 |
AstrocyteLayer(d_model) for _ in range(n_layers)
|
231 |
])
|
232 |
|
233 |
-
self.
|
|
|
|
|
234 |
|
235 |
-
#
|
236 |
-
self.
|
237 |
|
238 |
-
# Layer
|
239 |
-
self.
|
240 |
nn.LayerNorm(d_model) for _ in range(n_layers)
|
241 |
])
|
242 |
|
243 |
-
|
244 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
|
246 |
-
self.dropout = nn.Dropout(0.1)
|
247 |
self.classifier = None
|
248 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
249 |
def forward(self, x, edge_index, batch=None):
|
|
|
|
|
|
|
|
|
|
|
250 |
# Input processing
|
251 |
-
h = self.input_norm(self.input_proj(x))
|
252 |
|
253 |
-
#
|
|
|
|
|
|
|
254 |
for i in range(len(self.gcn_layers)):
|
255 |
gcn = self.gcn_layers[i]
|
256 |
-
astrocyte = self.astrocyte_layers[i]
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
h_gcn = self.dropout(h_gcn)
|
261 |
|
262 |
-
# Path
|
263 |
-
|
264 |
|
265 |
-
# Path
|
266 |
-
|
267 |
|
268 |
-
#
|
269 |
-
|
270 |
-
h_combined = h_combined.permute(1, 0, 2) # (nodes, 3, features)
|
271 |
-
h_momentum = self.global_momentum(h_combined.unsqueeze(0)).squeeze(0) # (nodes, 3, features)
|
272 |
-
h_momentum = h_momentum.mean(dim=1) # (nodes, features)
|
273 |
|
274 |
-
#
|
275 |
-
|
276 |
-
|
|
|
277 |
|
278 |
-
# Residual
|
279 |
-
h = norm(h + h_fused)
|
|
|
|
|
|
|
280 |
|
281 |
return h
|
282 |
|
283 |
def _init_classifier(self, num_classes, device):
|
284 |
if self.classifier is None:
|
285 |
self.classifier = nn.Sequential(
|
286 |
-
nn.Dropout(0.
|
287 |
nn.Linear(self.config['model']['d_model'], num_classes)
|
288 |
).to(device)
|
|
|
289 |
|
290 |
def get_performance_stats(self):
|
291 |
total_params = sum(p.numel() for p in self.parameters())
|
|
|
|
|
292 |
return {
|
293 |
'total_params': total_params,
|
|
|
294 |
'device': next(self.parameters()).device,
|
295 |
'dtype': next(self.parameters()).dtype,
|
296 |
'model_size': f"{total_params/1000:.1f}K parameters"
|
297 |
}
|
298 |
|
299 |
-
class
|
300 |
-
"""
|
301 |
def __init__(self, config):
|
302 |
super().__init__()
|
303 |
-
self.cognitive_mamba = CognitiveMambaGraphMamba(config)
|
304 |
self.config = config
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
305 |
self.classifier = None
|
306 |
|
307 |
def forward(self, x, edge_index, batch=None):
|
308 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
|
310 |
def _init_classifier(self, num_classes, device):
|
311 |
-
self.classifier
|
312 |
-
nn.
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
return self.classifier
|
317 |
|
318 |
def get_performance_stats(self):
|
319 |
-
|
|
|
|
|
|
|
|
|
|
|
320 |
|
321 |
def create_astrocyte_config():
|
322 |
-
"""
|
323 |
return {
|
324 |
'model': {
|
325 |
-
'd_model':
|
326 |
'd_state': 8,
|
327 |
'd_conv': 4,
|
328 |
'expand': 2,
|
329 |
-
'n_layers':
|
330 |
-
'dropout': 0.
|
331 |
},
|
332 |
'data': {
|
333 |
'batch_size': 1,
|
334 |
'test_split': 0.2
|
335 |
},
|
336 |
'training': {
|
337 |
-
'learning_rate': 0.
|
338 |
-
'weight_decay': 0.
|
339 |
-
'epochs':
|
340 |
-
'patience':
|
341 |
-
'warmup_epochs':
|
342 |
-
'min_lr': 1e-
|
343 |
'label_smoothing': 0.0,
|
344 |
-
'max_gap': 0.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
345 |
},
|
346 |
'ordering': {
|
347 |
'strategy': 'none',
|
@@ -350,9 +431,8 @@ def create_astrocyte_config():
|
|
350 |
'input_dim': 1433
|
351 |
}
|
352 |
|
353 |
-
#
|
354 |
-
|
355 |
-
|
356 |
-
HybridGraphMamba =
|
357 |
-
QuantumEnhancedGraphMamba =
|
358 |
-
create_regularized_config = create_astrocyte_config
|
|
|
8 |
|
9 |
logger = logging.getLogger(__name__)
|
10 |
|
11 |
+
class GraphDataAugmentation:
|
12 |
+
"""Enhanced data augmentation for overfitting prevention"""
|
13 |
+
@staticmethod
|
14 |
+
def augment_features(x, noise_level=0.1, dropout_prob=0.05):
|
15 |
+
if not torch.is_tensor(x) or x.size(0) == 0:
|
16 |
+
return x
|
17 |
+
# Feature noise
|
18 |
+
noise = torch.randn_like(x) * noise_level
|
19 |
+
x_aug = x + noise
|
20 |
+
# Feature masking
|
21 |
+
mask = torch.rand(x.shape, device=x.device) > dropout_prob
|
22 |
+
return x_aug * mask.float()
|
23 |
+
|
24 |
+
@staticmethod
|
25 |
+
def augment_edges(edge_index, drop_prob=0.1):
|
26 |
+
if not torch.is_tensor(edge_index) or edge_index.size(1) == 0:
|
27 |
+
return edge_index
|
28 |
+
edge_mask = torch.rand(edge_index.size(1), device=edge_index.device) > drop_prob
|
29 |
+
return edge_index[:, edge_mask]
|
30 |
+
|
31 |
+
class SimpleMambaBlock(nn.Module):
|
32 |
+
"""Simplified Mamba block that actually works"""
|
33 |
+
def __init__(self, d_model, d_state=16):
|
34 |
super().__init__()
|
35 |
self.d_model = d_model
|
36 |
+
self.d_state = d_state
|
37 |
+
self.d_inner = d_model * 2
|
38 |
|
39 |
+
# Core projections
|
40 |
+
self.in_proj = nn.Linear(d_model, self.d_inner * 2, bias=False)
|
41 |
+
self.conv1d = nn.Conv1d(self.d_inner, self.d_inner, 3, groups=self.d_inner, padding=1)
|
42 |
+
self.out_proj = nn.Linear(self.d_inner, d_model, bias=False)
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
+
# State space parameters
|
45 |
+
self.dt_proj = nn.Linear(self.d_inner, self.d_inner, bias=True)
|
46 |
+
self.B_proj = nn.Linear(self.d_inner, d_state, bias=False)
|
47 |
+
self.C_proj = nn.Linear(self.d_inner, d_state, bias=False)
|
48 |
|
49 |
+
# Initialize A matrix
|
50 |
+
A = torch.arange(1, d_state + 1, dtype=torch.float32)
|
51 |
+
A = A.unsqueeze(0).repeat(self.d_inner, 1)
|
52 |
+
self.A_log = nn.Parameter(torch.log(A))
|
53 |
+
self.D = nn.Parameter(torch.ones(self.d_inner))
|
54 |
|
55 |
+
self.dropout = nn.Dropout(0.1)
|
|
|
|
|
56 |
|
57 |
+
def forward(self, x):
|
58 |
+
batch_size, seq_len, d_model = x.shape
|
59 |
|
60 |
+
# Dual path
|
61 |
+
xz = self.in_proj(x) # (B, L, 2*d_inner)
|
62 |
+
x_inner, z = xz.chunk(2, dim=-1) # Each: (B, L, d_inner)
|
63 |
|
64 |
+
# Convolution
|
65 |
+
x_conv = x_inner.transpose(1, 2) # (B, d_inner, L)
|
66 |
+
x_conv = self.conv1d(x_conv) # (B, d_inner, L)
|
67 |
+
x_conv = x_conv.transpose(1, 2) # (B, L, d_inner)
|
68 |
+
x_conv = F.silu(x_conv)
|
69 |
|
70 |
+
# State space
|
71 |
+
y = self.selective_scan(x_conv)
|
72 |
|
73 |
+
# Gate and output
|
74 |
+
y = y * F.silu(z)
|
75 |
+
output = self.out_proj(y)
|
76 |
|
77 |
+
return self.dropout(output)
|
78 |
|
79 |
+
def selective_scan(self, x):
|
80 |
+
"""Simplified selective scan"""
|
81 |
+
batch_size, seq_len, d_inner = x.shape
|
|
|
|
|
82 |
|
83 |
+
# Get parameters
|
84 |
+
dt = F.softplus(self.dt_proj(x)) # (B, L, d_inner)
|
85 |
+
B = self.B_proj(x) # (B, L, d_state)
|
86 |
+
C = self.C_proj(x) # (B, L, d_state)
|
87 |
|
88 |
+
# Discretize A
|
89 |
+
A = -torch.exp(self.A_log) # (d_inner, d_state)
|
90 |
+
deltaA = torch.exp(dt.unsqueeze(-1) * A.unsqueeze(0).unsqueeze(0)) # (B, L, d_inner, d_state)
|
91 |
+
deltaB = dt.unsqueeze(-1) * B.unsqueeze(2) # (B, L, d_inner, d_state)
|
92 |
|
93 |
+
# Initialize state
|
94 |
+
h = torch.zeros(batch_size, d_inner, self.d_state, device=x.device)
|
95 |
+
outputs = []
|
96 |
|
97 |
+
# Sequential processing
|
98 |
+
for i in range(seq_len):
|
99 |
+
h = deltaA[:, i] * h + deltaB[:, i] * x[:, i].unsqueeze(-1)
|
100 |
+
y = torch.sum(h * C[:, i].unsqueeze(1), dim=-1) + self.D * x[:, i]
|
101 |
+
outputs.append(y)
|
102 |
|
103 |
+
return torch.stack(outputs, dim=1)
|
104 |
|
105 |
+
class CognitiveMomentumEngine(nn.Module):
|
106 |
+
"""Simplified cognitive momentum"""
|
107 |
+
def __init__(self, d_model):
|
108 |
super().__init__()
|
109 |
self.d_model = d_model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
110 |
|
111 |
+
# Momentum projections
|
112 |
+
self.momentum_proj = nn.Linear(d_model, d_model)
|
113 |
+
self.force_proj = nn.Linear(d_model, d_model)
|
114 |
|
115 |
+
# Memory
|
116 |
+
self.register_buffer('momentum_state', torch.zeros(d_model))
|
117 |
+
self.decay = 0.95
|
|
|
|
|
|
|
|
|
118 |
|
119 |
def forward(self, x):
|
|
|
120 |
if x.dim() == 2:
|
121 |
+
batch_size, d_model = x.shape
|
122 |
+
# Global momentum update
|
123 |
+
force = self.force_proj(x.mean(dim=0))
|
124 |
+
self.momentum_state = self.decay * self.momentum_state + (1 - self.decay) * force
|
125 |
+
|
126 |
+
# Apply momentum
|
127 |
+
momentum_effect = self.momentum_proj(self.momentum_state).unsqueeze(0).expand(batch_size, -1)
|
128 |
+
return x + momentum_effect * 0.1
|
129 |
+
else:
|
130 |
+
return x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
|
132 |
+
class AstrocyteLayer(nn.Module):
|
133 |
+
"""Simplified astrocyte processing"""
|
134 |
+
def __init__(self, d_model):
|
135 |
super().__init__()
|
136 |
self.d_model = d_model
|
137 |
+
self.d_astrocyte = d_model
|
|
|
138 |
|
139 |
+
# Fast pathway
|
140 |
+
self.fast_proj = nn.Linear(d_model, d_model)
|
141 |
+
self.fast_dropout = nn.Dropout(0.1)
|
|
|
142 |
|
143 |
+
# Slow pathway
|
144 |
+
self.slow_proj = nn.Linear(d_model, self.d_astrocyte)
|
145 |
+
self.slow_integrate = nn.Linear(self.d_astrocyte, d_model)
|
146 |
+
self.slow_dropout = nn.Dropout(0.1)
|
|
|
|
|
|
|
|
|
147 |
|
148 |
+
# Gating
|
149 |
+
self.gate = nn.Linear(d_model * 2, d_model)
|
|
|
150 |
|
151 |
+
# Memory
|
152 |
+
self.register_buffer('slow_memory', torch.zeros(self.d_astrocyte))
|
153 |
+
self.memory_decay = 0.9
|
154 |
|
155 |
+
def forward(self, x):
|
156 |
+
if x.dim() == 3:
|
157 |
+
x = x.squeeze(0)
|
158 |
|
159 |
+
batch_size = x.size(0)
|
|
|
160 |
|
161 |
+
# Fast processing
|
162 |
+
fast_out = self.fast_dropout(F.relu(self.fast_proj(x)))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
163 |
|
164 |
+
# Slow processing with memory
|
165 |
+
slow_input = self.slow_proj(x.mean(dim=0))
|
166 |
+
self.slow_memory = self.memory_decay * self.slow_memory + (1 - self.memory_decay) * slow_input
|
167 |
+
slow_out = self.slow_dropout(F.relu(self.slow_integrate(self.slow_memory)))
|
168 |
+
slow_out = slow_out.unsqueeze(0).expand(batch_size, -1)
|
|
|
|
|
|
|
|
|
|
|
169 |
|
170 |
+
# Combine
|
171 |
+
combined = torch.cat([fast_out, slow_out], dim=-1)
|
172 |
+
gated = torch.sigmoid(self.gate(combined))
|
173 |
|
174 |
+
return fast_out * gated + slow_out * (1 - gated)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
|
176 |
+
class RevolutionaryGraphMamba(nn.Module):
|
177 |
+
"""Complete revolutionary implementation"""
|
178 |
def __init__(self, config):
|
179 |
super().__init__()
|
180 |
|
|
|
186 |
# Input processing
|
187 |
self.input_proj = nn.Linear(input_dim, d_model)
|
188 |
self.input_norm = nn.LayerNorm(d_model)
|
189 |
+
self.input_dropout = nn.Dropout(0.2)
|
190 |
|
191 |
+
# Data augmentation
|
192 |
+
self.augmentation = GraphDataAugmentation()
|
193 |
+
|
194 |
+
# Core components
|
195 |
self.gcn_layers = nn.ModuleList([
|
196 |
GCNConv(d_model, d_model) for _ in range(n_layers)
|
197 |
])
|
198 |
|
|
|
199 |
self.astrocyte_layers = nn.ModuleList([
|
200 |
AstrocyteLayer(d_model) for _ in range(n_layers)
|
201 |
])
|
202 |
|
203 |
+
self.mamba_blocks = nn.ModuleList([
|
204 |
+
SimpleMambaBlock(d_model) for _ in range(n_layers)
|
205 |
+
])
|
206 |
|
207 |
+
# Cognitive momentum
|
208 |
+
self.momentum_engine = CognitiveMomentumEngine(d_model)
|
209 |
|
210 |
+
# Layer processing
|
211 |
+
self.layer_norms = nn.ModuleList([
|
212 |
nn.LayerNorm(d_model) for _ in range(n_layers)
|
213 |
])
|
214 |
|
215 |
+
self.layer_dropouts = nn.ModuleList([
|
216 |
+
nn.Dropout(0.1) for _ in range(n_layers)
|
217 |
+
])
|
218 |
+
|
219 |
+
# Fusion
|
220 |
+
self.fusion_weights = nn.Parameter(torch.tensor([0.4, 0.3, 0.3]))
|
221 |
+
self.fusion_proj = nn.Linear(d_model * 3, d_model)
|
222 |
+
|
223 |
+
# Output
|
224 |
+
self.output_proj = nn.Linear(d_model, d_model)
|
225 |
+
self.output_dropout = nn.Dropout(0.2)
|
226 |
|
|
|
227 |
self.classifier = None
|
228 |
|
229 |
+
# Initialize weights
|
230 |
+
self.apply(self._init_weights)
|
231 |
+
|
232 |
+
def _init_weights(self, module):
|
233 |
+
if isinstance(module, nn.Linear):
|
234 |
+
torch.nn.init.xavier_uniform_(module.weight)
|
235 |
+
if module.bias is not None:
|
236 |
+
torch.nn.init.zeros_(module.bias)
|
237 |
+
elif isinstance(module, nn.LayerNorm):
|
238 |
+
torch.nn.init.ones_(module.weight)
|
239 |
+
torch.nn.init.zeros_(module.bias)
|
240 |
+
|
241 |
def forward(self, x, edge_index, batch=None):
|
242 |
+
# Apply data augmentation during training
|
243 |
+
if self.training:
|
244 |
+
x = self.augmentation.augment_features(x)
|
245 |
+
edge_index = self.augmentation.augment_edges(edge_index)
|
246 |
+
|
247 |
# Input processing
|
248 |
+
h = self.input_dropout(self.input_norm(self.input_proj(x)))
|
249 |
|
250 |
+
# Apply cognitive momentum
|
251 |
+
h = self.momentum_engine(h)
|
252 |
+
|
253 |
+
# Multi-path processing
|
254 |
for i in range(len(self.gcn_layers)):
|
255 |
gcn = self.gcn_layers[i]
|
256 |
+
astrocyte = self.astrocyte_layers[i]
|
257 |
+
mamba = self.mamba_blocks[i]
|
258 |
+
norm = self.layer_norms[i]
|
259 |
+
dropout = self.layer_dropouts[i]
|
|
|
260 |
|
261 |
+
# Path 1: GCN (structural)
|
262 |
+
h_gcn = F.relu(gcn(h, edge_index))
|
263 |
|
264 |
+
# Path 2: Astrocyte (temporal)
|
265 |
+
h_astrocyte = astrocyte(h)
|
266 |
|
267 |
+
# Path 3: Mamba (sequential)
|
268 |
+
h_mamba = mamba(h.unsqueeze(0)).squeeze(0)
|
|
|
|
|
|
|
269 |
|
270 |
+
# Fusion
|
271 |
+
h_paths = torch.stack([h_gcn, h_astrocyte, h_mamba], dim=-1) # (nodes, d_model, 3)
|
272 |
+
weights = F.softmax(self.fusion_weights, dim=0) # (3,)
|
273 |
+
h_fused = torch.sum(h_paths * weights, dim=-1) # (nodes, d_model)
|
274 |
|
275 |
+
# Residual connection
|
276 |
+
h = dropout(norm(h + h_fused))
|
277 |
+
|
278 |
+
# Output processing
|
279 |
+
h = self.output_dropout(self.output_proj(h))
|
280 |
|
281 |
return h
|
282 |
|
283 |
def _init_classifier(self, num_classes, device):
|
284 |
if self.classifier is None:
|
285 |
self.classifier = nn.Sequential(
|
286 |
+
nn.Dropout(0.3),
|
287 |
nn.Linear(self.config['model']['d_model'], num_classes)
|
288 |
).to(device)
|
289 |
+
return self.classifier
|
290 |
|
291 |
def get_performance_stats(self):
|
292 |
total_params = sum(p.numel() for p in self.parameters())
|
293 |
+
trainable_params = sum(p.numel() for p in self.parameters() if p.requires_grad)
|
294 |
+
|
295 |
return {
|
296 |
'total_params': total_params,
|
297 |
+
'trainable_params': trainable_params,
|
298 |
'device': next(self.parameters()).device,
|
299 |
'dtype': next(self.parameters()).dtype,
|
300 |
'model_size': f"{total_params/1000:.1f}K parameters"
|
301 |
}
|
302 |
|
303 |
+
class SimpleGraphMamba(nn.Module):
|
304 |
+
"""Simplified but working version"""
|
305 |
def __init__(self, config):
|
306 |
super().__init__()
|
|
|
307 |
self.config = config
|
308 |
+
d_model = config['model']['d_model']
|
309 |
+
n_layers = config['model']['n_layers']
|
310 |
+
input_dim = config.get('input_dim', 1433)
|
311 |
+
|
312 |
+
# Simple architecture
|
313 |
+
self.input_proj = nn.Linear(input_dim, d_model)
|
314 |
+
self.input_norm = nn.LayerNorm(d_model)
|
315 |
+
|
316 |
+
# GCN backbone
|
317 |
+
self.gcn_layers = nn.ModuleList([
|
318 |
+
GCNConv(d_model, d_model) for _ in range(n_layers)
|
319 |
+
])
|
320 |
+
|
321 |
+
# Enhanced features
|
322 |
+
self.enhancements = nn.ModuleList([
|
323 |
+
nn.Sequential(
|
324 |
+
nn.Linear(d_model, d_model * 2),
|
325 |
+
nn.ReLU(),
|
326 |
+
nn.Dropout(0.1),
|
327 |
+
nn.Linear(d_model * 2, d_model)
|
328 |
+
) for _ in range(n_layers)
|
329 |
+
])
|
330 |
+
|
331 |
+
self.layer_norms = nn.ModuleList([
|
332 |
+
nn.LayerNorm(d_model) for _ in range(n_layers)
|
333 |
+
])
|
334 |
+
|
335 |
+
self.dropout = nn.Dropout(0.2)
|
336 |
self.classifier = None
|
337 |
|
338 |
def forward(self, x, edge_index, batch=None):
|
339 |
+
h = self.input_norm(self.input_proj(x))
|
340 |
+
|
341 |
+
for i, (gcn, enhance, norm) in enumerate(zip(self.gcn_layers, self.enhancements, self.layer_norms)):
|
342 |
+
# GCN processing
|
343 |
+
h_gcn = F.relu(gcn(h, edge_index))
|
344 |
+
|
345 |
+
# Enhancement
|
346 |
+
h_enhanced = enhance(h_gcn)
|
347 |
+
|
348 |
+
# Residual + norm
|
349 |
+
h = norm(h + h_enhanced)
|
350 |
+
h = self.dropout(h)
|
351 |
+
|
352 |
+
return h
|
353 |
|
354 |
def _init_classifier(self, num_classes, device):
|
355 |
+
if self.classifier is None:
|
356 |
+
self.classifier = nn.Sequential(
|
357 |
+
nn.Dropout(0.3),
|
358 |
+
nn.Linear(self.config['model']['d_model'], num_classes)
|
359 |
+
).to(device)
|
360 |
return self.classifier
|
361 |
|
362 |
def get_performance_stats(self):
|
363 |
+
total_params = sum(p.numel() for p in self.parameters())
|
364 |
+
return {
|
365 |
+
'total_params': total_params,
|
366 |
+
'device': next(self.parameters()).device,
|
367 |
+
'model_size': f"{total_params/1000:.1f}K parameters"
|
368 |
+
}
|
369 |
|
370 |
def create_astrocyte_config():
|
371 |
+
"""Optimized configuration"""
|
372 |
return {
|
373 |
'model': {
|
374 |
+
'd_model': 64, # Reduced to prevent overfitting
|
375 |
'd_state': 8,
|
376 |
'd_conv': 4,
|
377 |
'expand': 2,
|
378 |
+
'n_layers': 2, # Reduced layers
|
379 |
+
'dropout': 0.2
|
380 |
},
|
381 |
'data': {
|
382 |
'batch_size': 1,
|
383 |
'test_split': 0.2
|
384 |
},
|
385 |
'training': {
|
386 |
+
'learning_rate': 0.01,
|
387 |
+
'weight_decay': 0.005,
|
388 |
+
'epochs': 200,
|
389 |
+
'patience': 30,
|
390 |
+
'warmup_epochs': 10,
|
391 |
+
'min_lr': 1e-5,
|
392 |
'label_smoothing': 0.0,
|
393 |
+
'max_gap': 0.15
|
394 |
+
},
|
395 |
+
'ordering': {
|
396 |
+
'strategy': 'none',
|
397 |
+
'preserve_locality': True
|
398 |
+
},
|
399 |
+
'input_dim': 1433
|
400 |
+
}
|
401 |
+
|
402 |
+
def create_regularized_config():
|
403 |
+
"""Heavily regularized config for small datasets"""
|
404 |
+
return {
|
405 |
+
'model': {
|
406 |
+
'd_model': 32, # Very small
|
407 |
+
'd_state': 4,
|
408 |
+
'd_conv': 4,
|
409 |
+
'expand': 2,
|
410 |
+
'n_layers': 2,
|
411 |
+
'dropout': 0.3
|
412 |
+
},
|
413 |
+
'data': {
|
414 |
+
'batch_size': 1,
|
415 |
+
'test_split': 0.2
|
416 |
+
},
|
417 |
+
'training': {
|
418 |
+
'learning_rate': 0.005,
|
419 |
+
'weight_decay': 0.01,
|
420 |
+
'epochs': 150,
|
421 |
+
'patience': 20,
|
422 |
+
'warmup_epochs': 5,
|
423 |
+
'min_lr': 1e-6,
|
424 |
+
'label_smoothing': 0.1,
|
425 |
+
'max_gap': 0.1
|
426 |
},
|
427 |
'ordering': {
|
428 |
'strategy': 'none',
|
|
|
431 |
'input_dim': 1433
|
432 |
}
|
433 |
|
434 |
+
# Model aliases
|
435 |
+
GraphMamba = RevolutionaryGraphMamba
|
436 |
+
AstrocyteGraphMamba = RevolutionaryGraphMamba
|
437 |
+
HybridGraphMamba = SimpleGraphMamba # Fallback to simple version
|
438 |
+
QuantumEnhancedGraphMamba = SimpleGraphMamba
|
|