File size: 10,905 Bytes
f130c4a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
"""

Hierarchical Architectural Style Classifier

Combines global, local, and relationship modeling for architectural style classification.

"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Dict, List, Tuple, Optional
import timm
from transformers import ViTModel, ViTConfig


class GlobalStyleBranch(nn.Module):
    """Global branch for overall architectural composition."""
    
    def __init__(self, model_name: str = 'efficientnet_b4', num_classes: int = 25):
        super().__init__()
        self.backbone = timm.create_model(model_name, pretrained=True, num_classes=0)
        self.global_pool = nn.AdaptiveAvgPool2d(1)
        self.classifier = nn.Linear(self.backbone.num_features, num_classes)
        
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        features = self.backbone.forward_features(x)
        if isinstance(features, tuple):
            features = features[0]
        pooled = self.global_pool(features).flatten(1)
        return self.classifier(pooled)


class LocalDetailBranch(nn.Module):
    """Local branch for architectural elements using Vision Transformer."""
    
    def __init__(self, image_size: int = 224, patch_size: int = 16, 

                 num_classes: int = 25, dim: int = 768, depth: int = 12, 

                 heads: int = 12):
        super().__init__()
        self.vit_config = ViTConfig(
            image_size=image_size,
            patch_size=patch_size,
            num_classes=num_classes,
            hidden_size=dim,
            num_hidden_layers=depth,
            num_attention_heads=heads,
            intermediate_size=dim * 4
        )
        self.vit = ViTModel(self.vit_config)
        self.classifier = nn.Linear(dim, num_classes)
        
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        outputs = self.vit(x)
        pooled_output = outputs.pooler_output
        return self.classifier(pooled_output)


class RelationshipBranch(nn.Module):
    """Graph Neural Network for modeling architectural component relationships."""
    
    def __init__(self, num_classes: int = 25, hidden_dim: int = 256, 

                 num_layers: int = 3):
        super().__init__()
        # Use a simple CNN to extract features first
        self.feature_extractor = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(64, 128, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(128, 256, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.AdaptiveAvgPool2d((1, 1))
        )
        
        # Flatten and project to hidden dimension
        self.input_projection = nn.Linear(256, hidden_dim)
        self.gnn_layers = nn.ModuleList([
            nn.Linear(hidden_dim, hidden_dim) for _ in range(num_layers)
        ])
        self.classifier = nn.Linear(hidden_dim, num_classes)
        
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # Extract features using CNN
        features = self.feature_extractor(x)
        features = features.view(features.size(0), -1)  # Flatten
        
        # Simplified GNN implementation
        # In practice, you'd use torch_geometric for proper GNN
        x = self.input_projection(features)
        for layer in self.gnn_layers:
            x = F.relu(layer(x))
        return self.classifier(x)  # No need for mean since it's already pooled


class FeatureFusion(nn.Module):
    """Fuses features from different branches."""
    
    def __init__(self, global_dim: int, local_dim: int, relationship_dim: int, 

                 fusion_dim: int = 512):
        super().__init__()
        self.global_projection = nn.Linear(global_dim, fusion_dim)
        self.local_projection = nn.Linear(local_dim, fusion_dim)
        self.relationship_projection = nn.Linear(relationship_dim, fusion_dim)
        self.fusion_layer = nn.Linear(fusion_dim * 3, fusion_dim)
        
    def forward(self, global_feat: torch.Tensor, local_feat: torch.Tensor, 

                relationship_feat: torch.Tensor) -> torch.Tensor:
        global_proj = self.global_projection(global_feat)
        local_proj = self.local_projection(local_feat)
        relationship_proj = self.relationship_projection(relationship_feat)
        
        combined = torch.cat([global_proj, local_proj, relationship_proj], dim=1)
        return self.fusion_layer(combined)


class HierarchicalClassifier(nn.Module):
    """Hierarchical classifier for broad categories and fine-grained styles."""
    
    def __init__(self, input_dim: int = 512, broad_classes: int = 5, 

                 fine_classes: int = 25):
        super().__init__()
        self.broad_classifier = nn.Linear(input_dim, broad_classes)
        self.fine_classifier = nn.Linear(input_dim, fine_classes)
        
        # Style hierarchy mapping (simplified)
        self.style_hierarchy = {
            0: [0, 1, 2, 3, 4],      # Ancient
            1: [5, 6, 7, 8, 9],      # Medieval  
            2: [10, 11, 12, 13, 14], # Renaissance
            3: [15, 16, 17, 18, 19], # Modern
            4: [20, 21, 22, 23, 24]  # Contemporary
        }
        
    def forward(self, features: torch.Tensor) -> Dict[str, torch.Tensor]:
        broad_logits = self.broad_classifier(features)
        fine_logits = self.fine_classifier(features)
        
        # Apply hierarchical constraints
        constrained_fine_logits = self.apply_hierarchical_constraints(
            broad_logits, fine_logits
        )
        
        return {
            'broad_logits': broad_logits,
            'fine_logits': constrained_fine_logits,
            'broad_probs': F.softmax(broad_logits, dim=1),
            'fine_probs': F.softmax(constrained_fine_logits, dim=1)
        }
    
    def apply_hierarchical_constraints(self, broad_logits: torch.Tensor, 

                                     fine_logits: torch.Tensor) -> torch.Tensor:
        """Apply hierarchical constraints to ensure consistency."""
        broad_probs = F.softmax(broad_logits, dim=1)
        constrained_fine = fine_logits.clone()
        
        # Mask fine-grained logits based on broad category predictions
        for i in range(broad_probs.shape[1]):
            mask = (broad_probs[:, i] < 0.1).unsqueeze(1)  # Low confidence in broad category
            for fine_idx in self.style_hierarchy[i]:
                constrained_fine[:, fine_idx] = torch.where(
                    mask.squeeze(), 
                    constrained_fine[:, fine_idx] - 10.0,  # Penalize
                    constrained_fine[:, fine_idx]
                )
        
        return constrained_fine


class HierarchicalArchitecturalClassifier(nn.Module):
    """Main hierarchical architectural style classifier."""
    
    def __init__(self, num_broad_classes: int = 5, num_fine_classes: int = 25,

                 image_size: int = 224, use_pretrained: bool = True):
        super().__init__()
        
        # Initialize branches
        self.global_branch = GlobalStyleBranch(num_classes=num_fine_classes)
        self.local_branch = LocalDetailBranch(num_classes=num_fine_classes)
        self.relationship_branch = RelationshipBranch(num_classes=num_fine_classes)
        
        # Feature fusion
        self.feature_fusion = FeatureFusion(
            global_dim=num_fine_classes,
            local_dim=num_fine_classes, 
            relationship_dim=num_fine_classes
        )
        
        # Hierarchical classifier
        self.hierarchical_classifier = HierarchicalClassifier(
            input_dim=512,
            broad_classes=num_broad_classes,
            fine_classes=num_fine_classes
        )
        
        # Multi-scale attention
        self.attention = MultiScaleAttention(
            global_dim=num_fine_classes,
            local_dim=num_fine_classes
        )
        
    def forward(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
        # Extract features from each branch
        global_features = self.global_branch(x)
        local_features = self.local_branch(x)
        relationship_features = self.relationship_branch(x)
        
        # Apply attention mechanism
        attended_global, attended_local = self.attention(
            global_features, local_features
        )
        
        # Fuse features
        fused_features = self.feature_fusion(
            attended_global, attended_local, relationship_features
        )
        
        # Hierarchical classification
        outputs = self.hierarchical_classifier(fused_features)
        
        # Add attention weights for interpretability
        outputs['attention_weights'] = self.attention.get_attention_weights()
        
        return outputs
    
    def get_style_hierarchy(self) -> Dict[int, List[int]]:
        """Get the style hierarchy mapping."""
        return self.hierarchical_classifier.style_hierarchy


class MultiScaleAttention(nn.Module):
    """Multi-scale attention mechanism for interpretability."""
    
    def __init__(self, global_dim: int, local_dim: int, attention_dim: int = 256):
        super().__init__()
        self.global_projection = nn.Linear(global_dim, attention_dim)
        self.local_projection = nn.Linear(local_dim, attention_dim)
        self.attention_weights = nn.Parameter(torch.randn(attention_dim, attention_dim))
        self.attention_weights_history = []
        
    def forward(self, global_features: torch.Tensor, 

                local_features: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        global_proj = self.global_projection(global_features)
        local_proj = self.local_projection(local_features)
        
        # Compute attention weights
        attention_scores = torch.matmul(global_proj, self.attention_weights)
        attention_scores = torch.matmul(attention_scores, local_proj.transpose(-2, -1))
        attention_weights = F.softmax(attention_scores, dim=-1)
        
        # Store attention weights for visualization
        self.attention_weights_history.append(attention_weights.detach())
        
        # Apply attention
        attended_global = torch.matmul(attention_weights, global_features)
        attended_local = torch.matmul(attention_weights.transpose(-2, -1), local_features)
        
        return attended_global, attended_local
    
    def get_attention_weights(self) -> torch.Tensor:
        """Get the latest attention weights for visualization."""
        if self.attention_weights_history:
            return self.attention_weights_history[-1]
        return torch.zeros(1, 1, 1)  # Placeholder