File size: 7,570 Bytes
bc30c66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f270b46
bc30c66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models


# -------------------------------
# 1. SWIN WINDOW UTILS
# -------------------------------
def window_partition(x, window_size):
    """Partitions input tensor into windows of shape (B * num_windows, window_size*window_size, C)."""
    B, H, W, C = x.shape
    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous()
    windows = x.view(-1, window_size * window_size, C)
    return windows

def window_reverse(windows, window_size, H, W):
    """Reverses the window partition operation."""
    B = int(windows.shape[0] / (H * W / window_size / window_size))
    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous()
    x = x.view(B, H, W, -1)
    return x


# -------------------------------
# 2. SWIN WINDOW ATTENTION
# -------------------------------
class SwinWindowAttention(nn.Module):
    """Swin-style window attention block."""
    def __init__(self, embed_dim, window_size, num_heads, dropout=0.0):
        super(SwinWindowAttention, self).__init__()
        self.embed_dim = embed_dim
        self.window_size = window_size
        self.num_heads = num_heads
        self.mha = nn.MultiheadAttention(embed_dim, num_heads, dropout=dropout, batch_first=True)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        """Perform multi-head self-attn within windows."""
        B, C, H, W = x.shape
        x = x.permute(0, 2, 3, 1).contiguous()

        pad_h = (self.window_size - H % self.window_size) % self.window_size
        pad_w = (self.window_size - W % self.window_size) % self.window_size
        if pad_h or pad_w:
            x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
        Hp, Wp = x.shape[1], x.shape[2]

        windows = window_partition(x, self.window_size)  # (B*n_wins, win_size^2, C)

        attn_windows, _ = self.mha(windows, windows, windows)
        attn_windows = self.dropout(attn_windows)

        x = window_reverse(attn_windows, self.window_size, Hp, Wp)

        if pad_h or pad_w:
            x = x[:, :H, :W, :].contiguous()

        return x.permute(0, 3, 1, 2).contiguous()


# -------------------------------
# 3. GLAM
# -------------------------------
class GLAM(nn.Module):
    """Global-Local Attention Module (GLAM)."""
    def __init__(self, in_channels, reduction_ratio=8):
        super(GLAM, self).__init__()

        # Local Channel Attention
        self.local_channel_conv = nn.Conv2d(in_channels, in_channels // reduction_ratio, kernel_size=1)
        self.local_channel_act = nn.Sigmoid()
        self.local_channel_expand = nn.Conv2d(in_channels // reduction_ratio, in_channels, kernel_size=1)

        # Local Spatial Attention
        self.local_spatial_conv3 = nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=3, dilation=3)
        self.local_spatial_conv5 = nn.Conv2d(in_channels, in_channels, kernel_size=3, padding=5, dilation=5)
        self.local_spatial_merge = nn.Conv2d(in_channels * 3, in_channels, kernel_size=1)
        self.local_spatial_act = nn.Sigmoid()

        # Global Channel Attention
        self.global_avg_pool = nn.AdaptiveAvgPool2d(1)
        self.global_channel_fc1 = nn.Linear(in_channels, in_channels // reduction_ratio)
        self.global_channel_fc2 = nn.Linear(in_channels // reduction_ratio, in_channels)
        self.global_channel_act = nn.Sigmoid()

        # Global Spatial Attention
        self.global_spatial_conv = nn.Conv2d(in_channels, 1, kernel_size=1)
        self.global_spatial_softmax = nn.Softmax(dim=-1)

        # Weights
        self.local_attention_weight = nn.Parameter(torch.tensor(1.0))
        self.global_attention_weight = nn.Parameter(torch.tensor(1.0))

    def forward(self, x):
        # Local Channel Attention
        lca = self.local_channel_conv(x)
        lca = self.local_channel_act(lca)
        lca = self.local_channel_expand(lca)
        lca_out = lca * x

        # Local Spatial Attention
        lsa3 = self.local_spatial_conv3(x)
        lsa5 = self.local_spatial_conv5(x)
        lsa_cat = torch.cat([x, lsa3, lsa5], dim=1)
        lsa = self.local_spatial_merge(lsa_cat)
        lsa = self.local_spatial_act(lsa)
        lsa_out = lsa * lca_out
        lsa_out = lsa_out + lca_out

        # Global Channel Attention
        B, C, H, W = x.size()
        gca = self.global_avg_pool(x).view(B, C)
        gca = F.relu(self.global_channel_fc1(gca), inplace=True)
        gca = self.global_channel_fc2(gca)
        gca = self.global_channel_act(gca)
        gca = gca.view(B, C, 1, 1)
        gca_out = gca * x

        # Global Spatial Attention
        gsa = self.global_spatial_conv(x)      # [B, 1, H, W]
        gsa = gsa.view(B, -1)                  # [B, H*W]
        gsa = self.global_spatial_softmax(gsa)
        gsa = gsa.view(B, 1, H, W)
        gsa_out = gsa * gca_out
        gsa_out = gsa_out + gca_out

        # Final Fusion
        out = lsa_out * self.local_attention_weight + gsa_out * self.global_attention_weight + x
        return out


# -------------------------------
# 4. EFFICIENTNETB0_TRANSFORMERGLAM
# -------------------------------
class EfficientNetb0_TransformerGLAM(nn.Module):
    """EfficientNet-B0 + Swin-style Transformer + GLAM + Self-Adaptive Gating."""
    def __init__(self,
                 num_classes=3,
                 embed_dim=512,
                 num_heads=8,
                 mlp_dim=512,
                 dropout=0.5,
                 window_size=7,
                 reduction_ratio=8):
        super(EfficientNetb0_TransformerGLAM, self).__init__()

        # EfficientNet Backbone
        efficientnet = models.efficientnet_b0(weights=None)
        self.feature_extractor = nn.Sequential(*list(efficientnet.features.children()))
        self.conv1x1 = nn.Conv2d(1280, embed_dim, kernel_size=1)

        # Transformer path
        self.pre_attn_norm = nn.LayerNorm(embed_dim)
        self.swin_attn = SwinWindowAttention(embed_dim, window_size, num_heads, dropout)
        self.post_attn_norm = nn.LayerNorm(embed_dim)

        # GLAM path
        self.glam = GLAM(in_channels=embed_dim, reduction_ratio=reduction_ratio)

        # Self-adaptive gating
        self.gate_fc = nn.Linear(embed_dim, 1)

        # Final classification
        self.dropout = nn.Dropout(dropout)
        self.fc = nn.Linear(embed_dim, num_classes)

    def forward(self, x):
        # Backbone
        feats = self.feature_extractor(x)           # [B, 1280, H', W']
        feats = self.conv1x1(feats)                  # [B, embed_dim, H', W']

        B, C, H, W = feats.shape

        # Transformer path
        x_perm = feats.permute(0, 2, 3, 1).contiguous()
        x_norm = self.pre_attn_norm(x_perm)           # LN
        x_norm = x_norm.permute(0, 3, 1, 2).contiguous()
        x_norm = self.dropout(x_norm)

        T = self.swin_attn(x_norm)

        T_perm = T.permute(0, 2, 3, 1).contiguous()
        T_norm = self.post_attn_norm(T_perm)          # LN
        T_out = T_norm.permute(0, 3, 1, 2).contiguous()

        # GLAM path
        G_out = self.glam(feats)

        # Gating
        gap_feats = F.adaptive_avg_pool2d(feats, (1, 1)).view(B, C)
        g = torch.sigmoid(self.gate_fc(gap_feats))
        g = g.view(B, 1, 1, 1)

        # Final Fusion
        F_out = g * T_out + (1 - g) * G_out
        pooled = F.adaptive_avg_pool2d(F_out, (1, 1)).view(B, -1)

        return self.fc(pooled)