File size: 13,217 Bytes
da003cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316

import torch
import torch.nn as nn
import torch.nn.functional as F
import timm
import numpy as np

dog_breeds = ["Afghan_Hound", "African_Hunting_Dog", "Airedale", "American_Staffordshire_Terrier",
              "Appenzeller", "Australian_Terrier", "Bedlington_Terrier", "Bernese_Mountain_Dog", "Bichon_Frise",
              "Blenheim_Spaniel", "Border_Collie", "Border_Terrier", "Boston_Bull", "Bouvier_Des_Flandres",
              "Brabancon_Griffon", "Brittany_Spaniel", "Cardigan", "Chesapeake_Bay_Retriever",
              "Chihuahua", "Dachshund", "Dandie_Dinmont", "Doberman", "English_Foxhound", "English_Setter",
              "English_Springer", "EntleBucher", "Eskimo_Dog", "French_Bulldog", "German_Shepherd",
              "German_Short-Haired_Pointer", "Gordon_Setter", "Great_Dane", "Great_Pyrenees",
              "Greater_Swiss_Mountain_Dog","Havanese", "Ibizan_Hound", "Irish_Setter", "Irish_Terrier",
              "Irish_Water_Spaniel", "Irish_Wolfhound", "Italian_Greyhound", "Japanese_Spaniel",
              "Kerry_Blue_Terrier", "Labrador_Retriever", "Lakeland_Terrier", "Leonberg", "Lhasa",
              "Maltese_Dog", "Mexican_Hairless", "Newfoundland", "Norfolk_Terrier", "Norwegian_Elkhound",
              "Norwich_Terrier", "Old_English_Sheepdog", "Pekinese", "Pembroke", "Pomeranian",
              "Rhodesian_Ridgeback", "Rottweiler", "Saint_Bernard", "Saluki", "Samoyed",
              "Scotch_Terrier", "Scottish_Deerhound", "Sealyham_Terrier", "Shetland_Sheepdog", "Shiba_Inu",
              "Shih-Tzu", "Siberian_Husky", "Staffordshire_Bullterrier", "Sussex_Spaniel",
              "Tibetan_Mastiff", "Tibetan_Terrier", "Walker_Hound", "Weimaraner",
              "Welsh_Springer_Spaniel", "West_Highland_White_Terrier", "Yorkshire_Terrier",
              "Affenpinscher", "Basenji", "Basset", "Beagle", "Black-and-Tan_Coonhound", "Bloodhound",
              "Bluetick", "Borzoi", "Boxer", "Briard", "Bull_Mastiff", "Cairn", "Chow", "Clumber",
              "Cocker_Spaniel", "Collie", "Curly-Coated_Retriever", "Dhole", "Dingo",
              "Flat-Coated_Retriever", "Giant_Schnauzer", "Golden_Retriever", "Groenendael", "Keeshond",
              "Kelpie", "Komondor", "Kuvasz", "Malamute", "Malinois", "Miniature_Pinscher",
              "Miniature_Poodle", "Miniature_Schnauzer", "Otterhound", "Papillon", "Pug", "Redbone",
              "Schipperke", "Silky_Terrier", "Soft-Coated_Wheaten_Terrier", "Standard_Poodle",
              "Standard_Schnauzer", "Toy_Poodle", "Toy_Terrier", "Vizsla", "Whippet",
              "Wire-Haired_Fox_Terrier"]


class MorphologicalFeatureExtractor(nn.Module):

    def __init__(self, in_features):
        super().__init__()

        # 基礎特徵維度設置
        self.reduced_dim = in_features // 4
        self.spatial_size = max(7, int(np.sqrt(self.reduced_dim // 64)))

        # 1. 特徵空間轉換器:將一維特徵轉換為二維空間表示
        self.dimension_transformer = nn.Sequential(
            nn.Linear(in_features, self.spatial_size * self.spatial_size * 64),
            nn.LayerNorm(self.spatial_size * self.spatial_size * 64),
            nn.ReLU()
        )

        # 2. 形態特徵分析器:分析具體的形態特徵
        self.morphological_analyzers = nn.ModuleDict({
            # 體型分析器:分析整體比例和大小
            'body_proportion': nn.Sequential(
                # 使用大卷積核捕捉整體體型特徵
                nn.Conv2d(64, 128, kernel_size=7, padding=3),
                nn.BatchNorm2d(128),
                nn.ReLU(),
                # 使用較小的卷積核精煉特徵
                nn.Conv2d(128, 128, kernel_size=3, padding=1),
                nn.BatchNorm2d(128),
                nn.ReLU()
            ),

            # 頭部特徵分析器:關注耳朵、臉部等
            'head_features': nn.Sequential(
                # 中等大小的卷積核,適合分析頭部結構
                nn.Conv2d(64, 128, kernel_size=5, padding=2),
                nn.BatchNorm2d(128),
                nn.ReLU(),
                # 小卷積核捕捉細節
                nn.Conv2d(128, 128, kernel_size=3, padding=1),
                nn.BatchNorm2d(128),
                nn.ReLU()
            ),

            # 尾部特徵分析器
            'tail_features': nn.Sequential(
                nn.Conv2d(64, 128, kernel_size=5, padding=2),
                nn.BatchNorm2d(128),
                nn.ReLU(),
                nn.Conv2d(128, 128, kernel_size=3, padding=1),
                nn.BatchNorm2d(128),
                nn.ReLU()
            ),

            # 毛髮特徵分析器:分析毛髮長度、質地等
            'fur_features': nn.Sequential(
                # 使用多個小卷積核捕捉毛髮紋理
                nn.Conv2d(64, 128, kernel_size=3, padding=1),
                nn.BatchNorm2d(128),
                nn.ReLU(),
                nn.Conv2d(128, 128, kernel_size=3, padding=1),
                nn.BatchNorm2d(128),
                nn.ReLU()
            ),

            # 顏色特徵分析器:分析顏色分佈
            'color_pattern': nn.Sequential(
                # 第一層:捕捉基本顏色分布
                nn.Conv2d(64, 128, kernel_size=3, padding=1),
                nn.BatchNorm2d(128),
                nn.ReLU(),

                # 第二層:分析顏色模式和花紋
                nn.Conv2d(128, 128, kernel_size=3, padding=1),
                nn.BatchNorm2d(128),
                nn.ReLU(),

                # 第三層:整合顏色信息
                nn.Conv2d(128, 128, kernel_size=1),
                nn.BatchNorm2d(128),
                nn.ReLU()
            )
        })

        # 3. 特徵注意力機制:動態關注不同特徵
        self.feature_attention = nn.MultiheadAttention(
            embed_dim=128,
            num_heads=8,
            dropout=0.1,
            batch_first=True
        )

        # 4. 特徵關係分析器:分析不同特徵之間的關係
        self.relation_analyzer = nn.Sequential(
            nn.Linear(128 * 5, 256),  # 4個特徵分析器的輸出
            nn.LayerNorm(256),
            nn.ReLU(),
            nn.Linear(256, 128),
            nn.LayerNorm(128),
            nn.ReLU()
        )

        # 5. 特徵整合器:將所有特徵智能地組合在一起
        self.feature_integrator = nn.Sequential(
            nn.Linear(128 * 6, in_features),  # 5個原始特徵 + 1個關係特徵
            nn.LayerNorm(in_features),
            nn.ReLU()
        )

    def forward(self, x):
        batch_size = x.size(0)

        # 1. 將特徵轉換為空間形式
        spatial_features = self.dimension_transformer(x).view(
            batch_size, 64, self.spatial_size, self.spatial_size
        )

        # 2. 分析各種形態特徵
        morphological_features = {}
        for name, analyzer in self.morphological_analyzers.items():
            # 提取特定形態特徵
            features = analyzer(spatial_features)
            # 使用自適應池化統一特徵大小
            pooled_features = F.adaptive_avg_pool2d(features, (1, 1))
            # 重塑特徵為向量形式
            morphological_features[name] = pooled_features.view(batch_size, -1)

        # 3. 特徵注意力處理
        # 將所有特徵堆疊成序列
        stacked_features = torch.stack(list(morphological_features.values()), dim=1)
        # 應用注意力機制
        attended_features, _ = self.feature_attention(
            stacked_features, stacked_features, stacked_features
        )

        # 4. 分析特徵之間的關係
        # 將所有特徵連接起來
        combined_features = torch.cat(list(morphological_features.values()), dim=1)
        # 提取特徵間的關係
        relation_features = self.relation_analyzer(combined_features)

        # 5. 特徵整合
        # 將原始特徵和關係特徵結合
        final_features = torch.cat([
            *morphological_features.values(),
            relation_features
        ], dim=1)

        # 6. 最終整合
        integrated_features = self.feature_integrator(final_features)

        # 添加殘差連接
        return integrated_features + x


class MultiHeadAttention(nn.Module):

    def __init__(self, in_dim, num_heads=8):
        """
        Initializes the MultiHeadAttention module.
        Args:
            in_dim (int): Dimension of the input features.
            num_heads (int): Number of attention heads. Defaults to 8.
        """
        super().__init__()
        self.num_heads = num_heads
        self.head_dim = max(1, in_dim // num_heads)  
        self.scaled_dim = self.head_dim * num_heads  
        self.fc_in = nn.Linear(in_dim, self.scaled_dim)  
        self.query = nn.Linear(self.scaled_dim, self.scaled_dim)  # Query projection
        self.key = nn.Linear(self.scaled_dim, self.scaled_dim)  # Key projection
        self.value = nn.Linear(self.scaled_dim, self.scaled_dim)  # Value projection
        self.fc_out = nn.Linear(self.scaled_dim, in_dim)  # Linear layer to project output back to in_dim

    def forward(self, x):
        """
        Forward pass for multi-head attention mechanism.
        Args:
            x (Tensor): Input tensor of shape (batch_size, input_dim).
            x 是 (N,D), N:批次大小, D:輸入特徵維度
        Returns:
            Tensor: Output tensor after applying attention mechanism.
        """
        N = x.shape[0]  # Batch size
        x = self.fc_in(x)  # Project input to scaled_dim
        q = self.query(x).view(N, self.num_heads, self.head_dim)  # Compute queries
        k = self.key(x).view(N, self.num_heads, self.head_dim)  # Compute keys
        v = self.value(x).view(N, self.num_heads, self.head_dim)  # Compute values

        # Calculate attention scores
        energy = torch.einsum("nqd,nkd->nqk", [q, k])  
        attention = F.softmax(energy / (self.head_dim ** 0.5), dim=2)  # Apply softmax with scaling

        # Compute weighted sum of values based on attention scores
        out = torch.einsum("nqk,nvd->nqd", [attention, v]) 
        out = out.reshape(N, self.scaled_dim)  # Concatenate all heads
        out = self.fc_out(out)  # Project back to original input dimension
        return out


class BaseModel(nn.Module):

    def __init__(self, num_classes, device='cuda' if torch.cuda.is_available() else 'cpu'):
        super().__init__()
        self.device = device

        # 1. Initialize backbone
        self.backbone = timm.create_model(
                'convnextv2_base',
                pretrained=True,
                num_classes=0
        )

        # 2. 使用測試數據來確定實際的特徵維度
        with torch.no_grad():  
            dummy_input = torch.randn(1, 3, 224, 224)
            features = self.backbone(dummy_input)

            if len(features.shape) > 2:
                features = features.mean([-2, -1])

            self.feature_dim = features.shape[1]

        print(f"Feature Dimension from V2 backbone: {self.feature_dim}")

        # 3. Setup multi-head attention layer
        self.num_heads = max(1, min(8, self.feature_dim // 64))
        self.attention = MultiHeadAttention(self.feature_dim, num_heads=self.num_heads)

        # 4. Setup classifier
        self.classifier = nn.Sequential(
            nn.LayerNorm(self.feature_dim),
            nn.Dropout(0.3),
            nn.Linear(self.feature_dim, num_classes)
        )

        self.morphological_extractor = MorphologicalFeatureExtractor(
            in_features=self.feature_dim
        )

        self.feature_fusion = nn.Sequential(
            nn.Linear(self.feature_dim * 3, self.feature_dim),  
            nn.LayerNorm(self.feature_dim),
            nn.ReLU(),
            nn.Linear(self.feature_dim, self.feature_dim),
            nn.LayerNorm(self.feature_dim),
            nn.ReLU()
        )

    def forward(self, x):
        """
        Forward propagation process, combining V2's FCCA and multi-head attention mechanism
        Args:
            x (Tensor): Input image tensor of shape [batch_size, channels, height, width]
        Returns:
            Tuple[Tensor, Tensor]: Classification logits and attention features
        """
        x = x.to(self.device)

        # 1. Extract base features
        features = self.backbone(x)
        if len(features.shape) > 2:
            features = features.mean([-2, -1])

        # 2. Extract morphological features (including all detail features)
        morphological_features = self.morphological_extractor(features)

        # 3. Feature fusion (note dimension alignment with new fusion layer)
        combined_features = torch.cat([
            features,  # Original features
            morphological_features,  # Morphological features
            features * morphological_features  # Feature interaction information
        ], dim=1)
        fused_features = self.feature_fusion(combined_features)

        # 4. Apply attention mechanism
        attended_features = self.attention(fused_features)

        # 5. Final classifier
        logits = self.classifier(attended_features)

        return logits, attended_features