调整模型参数

gui/model.py

@@ -2,20 +2,20 @@ import torch
 import torch.nn as nn
 
 
-# 轻量双通道注意力
+class DualAttention(nn.Module):
-class LiteDualAttention(nn.Module):
     def __init__(self, in_channels):
         super().__init__()
+        middle_channels = max(in_channels // 32, 1)
         self.channel_attn = nn.Sequential(
             nn.AdaptiveAvgPool2d(1),
-            nn.Conv2d(in_channels, in_channels // 16, 1, bias=False),
+            nn.Conv2d(in_channels, middle_channels, 1, bias=False),
             nn.SiLU(),
-            nn.Conv2d(in_channels // 16, in_channels, 1, bias=False),
+            nn.Conv2d(middle_channels, in_channels, 1, bias=False),
             nn.Sigmoid()
         )
 
         self.spatial_attn = nn.Sequential(
-            nn.Conv2d(2, 1, 5, padding=2, bias=False),
+            nn.Conv2d(2, 1, 3, padding=1, bias=False),
             nn.Sigmoid()
         )
 
@@ -29,43 +29,77 @@ class LiteDualAttention(nn.Module):
             x.std(dim=1, keepdim=True)
         ], dim=1)
 
-        # 空间加权（5x5卷积生成单通道注意力）
+        # 空间加权
         sa = self.spatial_attn(stats) * ca
         return sa
 
 
-class CNNClassifier(nn.Module):
+class CustomCNN(nn.Module):
     def __init__(self, num_classes):
         super().__init__()
         self.features = nn.Sequential(
-            nn.Conv2d(3, 32, 3, stride=2, padding=1, bias=False),
+            nn.Conv2d(3, 24, 3, stride=2, padding=1, bias=False),
-            LiteDualAttention(32),
+            DualAttention(24),
-            nn.BatchNorm2d(32),
+            nn.BatchNorm2d(24),
             nn.Hardswish(inplace=True),
 
-            nn.Conv2d(32, 48, 3, stride=2, padding=1, bias=False),
+            nn.Conv2d(24, 36, 3, stride=2, padding=1, bias=False),
-            LiteDualAttention(48),
+            DualAttention(36),
+            nn.BatchNorm2d(36),
+            nn.Hardswish(inplace=True),
+
+            nn.Conv2d(36, 48, 3, stride=2, padding=1, bias=False),
+            DualAttention(48),
             nn.BatchNorm2d(48),
             nn.Hardswish(inplace=True),
 
-            nn.Conv2d(48, 64, 3, stride=2, padding=1, bias=False),
+            nn.Conv2d(48, 64, 3, padding=2, dilation=2, bias=False),
-            LiteDualAttention(64),
+            DualAttention(64),
             nn.BatchNorm2d(64),
-            nn.Hardswish(inplace=True),
-
-            nn.Conv2d(64, 96, 3, padding=2, dilation=2, bias=False),
-            LiteDualAttention(96),
-            nn.BatchNorm2d(96),
             nn.Hardswish(inplace=True)
         )
 
         self.classifier = nn.Sequential(
             nn.AdaptiveAvgPool2d(1),
             nn.Flatten(),
-            nn.Linear(96, 256),
+            nn.Linear(64, 192),
             nn.Hardswish(inplace=True),
             nn.Dropout(0.3),
-            nn.Linear(256, num_classes)
+            nn.Linear(192, num_classes)
+        )
+
+    def forward(self, x):
+        return self.classifier(self.features(x))
+
+
+class SimpleCNN(nn.Module):
+    def __init__(self, num_classes):
+        super().__init__()
+        self.features = nn.Sequential(
+            nn.Conv2d(3, 24, 3, stride=2, padding=1, bias=False),
+            nn.BatchNorm2d(24),
+            nn.Hardswish(inplace=True),
+
+            nn.Conv2d(24, 36, 3, stride=2, padding=1, bias=False),
+            nn.BatchNorm2d(36),
+            nn.Hardswish(inplace=True),
+
+            nn.Conv2d(36, 48, 3, stride=2, padding=1, bias=False),
+            nn.BatchNorm2d(48),
+            nn.Hardswish(inplace=True),
+
+            nn.Conv2d(48, 64, 3, padding=2, dilation=2, bias=False),
+            nn.BatchNorm2d(64),
+            nn.Hardswish(inplace=True)
+        )
+
+        self.classifier = nn.Sequential(
+            nn.AdaptiveAvgPool2d(1),
+            nn.Flatten(),
+            nn.Linear(64, 192),
+            nn.Hardswish(inplace=True),
+            nn.Dropout(0.3),
+            nn.Linear(192, num_classes)
         )
 
     def forward(self, x):