File size: 9,133 Bytes
c681cda
 
 
 
 
 
 
e4d5cc2
c681cda
 
 
 
 
 
 
 
 
 
 
 
e4d5cc2
c681cda
 
 
e4d5cc2
c681cda
e4d5cc2
c681cda
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4d5cc2
1bdb453
 
e4d5cc2
 
 
1bdb453
c681cda
 
e4d5cc2
 
 
 
 
 
1bdb453
 
 
 
e4d5cc2
 
 
 
 
 
 
c681cda
 
 
e4d5cc2
c681cda
 
 
1bdb453
 
 
 
c681cda
 
1bdb453
c681cda
 
 
 
 
 
 
 
 
 
 
 
 
 
e4d5cc2
c681cda
1bdb453
 
 
 
c681cda
e4d5cc2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c681cda
1bdb453
 
 
c681cda
 
1bdb453
 
 
c681cda
 
 
 
 
 
 
e4d5cc2
 
 
 
 
 
 
 
 
 
 
c681cda
 
 
 
 
 
 
 
 
 
 
e4d5cc2
1bdb453
 
 
 
c681cda
 
 
 
 
 
 
 
 
 
 
 
1bdb453
 
 
 
 
 
 
 
 
c681cda
 
 
 
 
 
 
 
 
 
 
 
 
 
e4d5cc2
c681cda
e4d5cc2
 
 
 
 
 
 
 
 
c681cda
e4d5cc2
1bdb453
 
 
 
 
 
 
 
e4d5cc2
 
 
 
 
 
 
 
 
 
1bdb453
 
e4d5cc2
 
1bdb453
e4d5cc2
 
 
 
 
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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
import torch
import torch.nn as nn
from .mamba_block import MambaBlock
from .graph_sequencer import GraphSequencer, PositionalEncoder

class GraphMamba(nn.Module):
    """
    Production Graph-Mamba model with training optimizations
    """
    
    def __init__(self, config):
        super().__init__()
        
        self.config = config
        self.d_model = config['model']['d_model']
        self.n_layers = config['model']['n_layers']
        self.dropout = config['model']['dropout']
        self.ordering_strategy = config['ordering']['strategy']
        
        # Input projection (dynamic input dimension)
        self.input_proj = None
        
        # Positional encoding
        self.pos_encoder = PositionalEncoder()
        self.pos_embed = nn.Linear(11, self.d_model)
        
        # Mamba layers with residual connections
        self.mamba_layers = nn.ModuleList([
            MambaBlock(
                d_model=self.d_model,
                d_state=config['model']['d_state'],
                d_conv=config['model']['d_conv'],
                expand=config['model']['expand']
            )
            for _ in range(self.n_layers)
        ])
        
        # Layer norms
        self.layer_norms = nn.ModuleList([
            nn.LayerNorm(self.d_model)
            for _ in range(self.n_layers)
        ])
        
        # Dropout
        self.dropout_layer = nn.Dropout(self.dropout)
        
        # Graph sequencer
        self.sequencer = GraphSequencer()
        
        # Classification head (initialized dynamically)
        self.classifier = None
        
        # Cache for efficiency
        self._cache = {}
        
    def _init_input_proj(self, input_dim, device):
        """Initialize input projection dynamically"""
        if self.input_proj is None:
            self.input_proj = nn.Sequential(
                nn.Linear(input_dim, self.d_model),
                nn.LayerNorm(self.d_model),
                nn.ReLU(),
                nn.Dropout(self.dropout * 0.5)
            ).to(device)
            
    def _init_classifier(self, num_classes, device):
        """Initialize classifier dynamically"""
        if self.classifier is None:
            self.classifier = nn.Sequential(
                nn.Linear(self.d_model, self.d_model // 2),
                nn.LayerNorm(self.d_model // 2),
                nn.ReLU(),
                nn.Dropout(self.dropout),
                nn.Linear(self.d_model // 2, num_classes)
            ).to(device)
            
    def forward(self, x, edge_index, batch=None):
        """
        Forward pass with training optimizations
        """
        num_nodes = x.size(0)
        input_dim = x.size(1)
        device = x.device
        
        # Move all components to correct device
        self.to(device)
        
        # Initialize input projection if needed
        self._init_input_proj(input_dim, device)
        
        # Project input features
        h = self.input_proj(x)  # (num_nodes, d_model)
        
        if batch is None:
            # Single graph processing
            h = self._process_single_graph(h, edge_index)
        else:
            # Batch processing
            h = self._process_batch(h, edge_index, batch)
            
        return h
    
    def _process_single_graph(self, h, edge_index):
        """Process a single graph with caching"""
        num_nodes = h.size(0)
        device = h.device
        
        # Ensure edge_index is on correct device
        edge_index = edge_index.to(device)
        
        # Cache key for ordering
        cache_key = f"{self.ordering_strategy}_{num_nodes}_{edge_index.shape[1]}"
        
        # Get ordering (with caching during training)
        if cache_key not in self._cache or not self.training:
            if self.ordering_strategy == "spectral":
                order = self.sequencer.spectral_ordering(edge_index, num_nodes)
            elif self.ordering_strategy == "degree":
                order = self.sequencer.degree_ordering(edge_index, num_nodes)
            elif self.ordering_strategy == "community":
                order = self.sequencer.community_ordering(edge_index, num_nodes)
            else:  # default to BFS
                order = self.sequencer.bfs_ordering(edge_index, num_nodes)
            
            if self.training:
                self._cache[cache_key] = order
        else:
            order = self._cache[cache_key]
        
        # Ensure order is on correct device
        order = order.to(device)
        
        # Add positional encoding
        seq_pos, distances = self.pos_encoder.encode_positions(h, edge_index, order)
        seq_pos = seq_pos.to(device)
        distances = distances.to(device)
        
        pos_features = torch.cat([seq_pos, distances], dim=1)  # (num_nodes, 11)
        pos_embed = self.pos_embed(pos_features)
        
        # Reorder nodes for sequential processing
        h_ordered = h[order] + pos_embed[order]  # Add positional encoding
        h_ordered = h_ordered.unsqueeze(0)  # (1, num_nodes, d_model)
        
        # Process through Mamba layers with residual connections
        for i, (mamba, ln) in enumerate(zip(self.mamba_layers, self.layer_norms)):
            # Pre-norm residual connection with gradient scaling
            residual = h_ordered
            h_ordered = ln(h_ordered)
            h_ordered = mamba(h_ordered)
            h_ordered = residual + self.dropout_layer(h_ordered)
            
            # Layer-wise learning rate scaling
            if self.training:
                h_ordered = h_ordered * (1.0 - 0.1 * i / self.n_layers)
        
        # Restore original order
        h_out = h_ordered.squeeze(0)  # (num_nodes, d_model)
        
        # Create inverse mapping
        inverse_order = torch.argsort(order)
        h_final = h_out[inverse_order]
        
        return h_final
    
    def _process_batch(self, h, edge_index, batch):
        """Process batched graphs efficiently"""
        device = h.device
        batch = batch.to(device)
        edge_index = edge_index.to(device)
        
        batch_size = batch.max().item() + 1
        outputs = []
        
        for b in range(batch_size):
            # Extract subgraph
            mask = batch == b
            batch_h = h[mask]
            
            # Get edges for this graph
            edge_mask = mask[edge_index[0]] & mask[edge_index[1]]
            batch_edges = edge_index[:, edge_mask]
            
            if batch_edges.shape[1] > 0:
                # Reindex edges to local indices
                node_indices = torch.where(mask)[0]
                node_map = torch.zeros(h.size(0), dtype=torch.long, device=device)
                node_map[node_indices] = torch.arange(batch_h.size(0), device=device)
                batch_edges_local = node_map[batch_edges]
            else:
                # Empty graph
                batch_edges_local = torch.empty((2, 0), dtype=torch.long, device=device)
            
            # Process subgraph
            batch_output = self._process_single_graph(batch_h, batch_edges_local)
            outputs.append(batch_output)
        
        # Reconstruct full batch
        h_out = torch.zeros_like(h)
        for b, output in enumerate(outputs):
            mask = batch == b
            h_out[mask] = output
            
        return h_out
    
    def get_graph_embedding(self, h, batch=None):
        """Get graph-level representation with multiple pooling"""
        if batch is None:
            # Single graph - multiple pooling strategies
            mean_pool = h.mean(dim=0, keepdim=True)
            max_pool = h.max(dim=0)[0].unsqueeze(0)
            
            # Attention pooling
            attn_weights = torch.softmax(h.sum(dim=1), dim=0)
            attn_pool = (h * attn_weights.unsqueeze(1)).sum(dim=0, keepdim=True)
            
            return torch.cat([mean_pool, max_pool, attn_pool], dim=1)
        else:
            # Batched graphs
            device = h.device
            batch = batch.to(device)
            batch_size = batch.max().item() + 1
            
            graph_embeddings = []
            for b in range(batch_size):
                mask = batch == b
                if mask.any():
                    batch_h = h[mask]
                    
                    # Multiple pooling for this graph
                    mean_pool = batch_h.mean(dim=0)
                    max_pool = batch_h.max(dim=0)[0]
                    
                    attn_weights = torch.softmax(batch_h.sum(dim=1), dim=0)
                    attn_pool = (batch_h * attn_weights.unsqueeze(1)).sum(dim=0)
                    
                    graph_emb = torch.cat([mean_pool, max_pool, attn_pool])
                    graph_embeddings.append(graph_emb)
                else:
                    # Empty graph
                    graph_embeddings.append(torch.zeros(h.size(1) * 3, device=device))
            
            return torch.stack(graph_embeddings)
    
    def clear_cache(self):
        """Clear ordering cache"""
        self._cache.clear()