注意力可視化

訓練模型

包含通道注意力模塊和CNN模型的定義（通道注意力的插入）

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
import numpy as np# 設置中文字體支持
plt.rcParams["font.family"] = ["SimHei"]
plt.rcParams['axes.unicode_minus'] = False  # 解決負號顯示問題# 檢查GPU是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"使用設備: {device}")# 1. 數據預處理
# 訓練集：使用多種數據增強方法提高模型泛化能力
train_transform = transforms.Compose([# 隨機裁剪圖像，從原圖中隨機截取32x32大小的區域transforms.RandomCrop(32, padding=4),# 隨機水平翻轉圖像（概率0.5）transforms.RandomHorizontalFlip(),# 隨機顏色抖動：亮度、對比度、飽和度和色調隨機變化transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),# 隨機旋轉圖像（最大角度15度）transforms.RandomRotation(15),# 將PIL圖像或numpy數組轉換為張量transforms.ToTensor(),# 標準化處理：每個通道的均值和標準差，使數據分布更合理transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])# 測試集：僅進行必要的標準化，保持數據原始特性，標準化不損失數據信息，可還原
test_transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])# 2. 加載CIFAR-10數據集
train_dataset = datasets.CIFAR10(root='./data',train=True,download=True,transform=train_transform  # 使用增強后的預處理
)test_dataset = datasets.CIFAR10(root='./data',train=False,transform=test_transform  # 測試集不使用增強
)# 3. 創建數據加載器
batch_size = 64
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)# ===================== 新增：通道注意力模塊（SE模塊） =====================
class ChannelAttention(nn.Module):"""通道注意力模塊(Squeeze-and-Excitation)"""def __init__(self, in_channels, reduction_ratio=16):"""參數:in_channels: 輸入特征圖的通道數reduction_ratio: 降維比例，用于減少參數量"""super(ChannelAttention, self).__init__()# 全局平均池化 - 將空間維度壓縮為1x1，保留通道信息self.avg_pool = nn.AdaptiveAvgPool2d(1)# 全連接層 + 激活函數，用于學習通道間的依賴關系self.fc = nn.Sequential(# 降維：壓縮通道數，減少計算量nn.Linear(in_channels, in_channels // reduction_ratio, bias=False),nn.ReLU(inplace=True),# 升維：恢復原始通道數nn.Linear(in_channels // reduction_ratio, in_channels, bias=False),# Sigmoid將輸出值歸一化到[0,1]，表示通道重要性權重nn.Sigmoid())def forward(self, x):"""參數:x: 輸入特征圖，形狀為 [batch_size, channels, height, width]返回:加權后的特征圖，形狀不變"""batch_size, channels, height, width = x.size()# 1. 全局平均池化：[batch_size, channels, height, width] → [batch_size, channels, 1, 1]avg_pool_output = self.avg_pool(x)# 2. 展平為一維向量：[batch_size, channels, 1, 1] → [batch_size, channels]avg_pool_output = avg_pool_output.view(batch_size, channels)# 3. 通過全連接層學習通道權重：[batch_size, channels] → [batch_size, channels]channel_weights = self.fc(avg_pool_output)# 4. 重塑為二維張量：[batch_size, channels] → [batch_size, channels, 1, 1]channel_weights = channel_weights.view(batch_size, channels, 1, 1)# 5. 將權重應用到原始特征圖上（逐通道相乘）return x * channel_weights  # 輸出形狀：[batch_size, channels, height, width]# 4. 定義CNN模型的定義（通道注意力的插入）
class CNN(nn.Module):def __init__(self):super(CNN, self).__init__()  # ---------------------- 第一個卷積塊 ----------------------self.conv1 = nn.Conv2d(3, 32, 3, padding=1)self.bn1 = nn.BatchNorm2d(32)self.relu1 = nn.ReLU()# 新增：插入通道注意力模塊（SE模塊）self.ca1 = ChannelAttention(in_channels=32, reduction_ratio=16)  self.pool1 = nn.MaxPool2d(2, 2)  # ---------------------- 第二個卷積塊 ----------------------self.conv2 = nn.Conv2d(32, 64, 3, padding=1)self.bn2 = nn.BatchNorm2d(64)self.relu2 = nn.ReLU()# 新增：插入通道注意力模塊（SE模塊）self.ca2 = ChannelAttention(in_channels=64, reduction_ratio=16)  self.pool2 = nn.MaxPool2d(2)  # ---------------------- 第三個卷積塊 ----------------------self.conv3 = nn.Conv2d(64, 128, 3, padding=1)self.bn3 = nn.BatchNorm2d(128)self.relu3 = nn.ReLU()# 新增：插入通道注意力模塊（SE模塊）self.ca3 = ChannelAttention(in_channels=128, reduction_ratio=16)  self.pool3 = nn.MaxPool2d(2)  # ---------------------- 全連接層（分類器） ----------------------self.fc1 = nn.Linear(128 * 4 * 4, 512)self.dropout = nn.Dropout(p=0.5)self.fc2 = nn.Linear(512, 10)def forward(self, x):# ---------- 卷積塊1處理 ----------x = self.conv1(x)       x = self.bn1(x)         x = self.relu1(x)       x = self.ca1(x)  # 應用通道注意力x = self.pool1(x)       # ---------- 卷積塊2處理 ----------x = self.conv2(x)       x = self.bn2(x)         x = self.relu2(x)       x = self.ca2(x)  # 應用通道注意力x = self.pool2(x)       # ---------- 卷積塊3處理 ----------x = self.conv3(x)       x = self.bn3(x)         x = self.relu3(x)       x = self.ca3(x)  # 應用通道注意力x = self.pool3(x)       # ---------- 展平與全連接層 ----------x = x.view(-1, 128 * 4 * 4)  x = self.fc1(x)           x = self.relu3(x)         x = self.dropout(x)       x = self.fc2(x)           return x  # 重新初始化模型，包含通道注意力模塊
model = CNN()
model = model.to(device)  # 將模型移至GPU（如果可用）criterion = nn.CrossEntropyLoss()  # 交叉熵損失函數
optimizer = optim.Adam(model.parameters(), lr=0.001)  # Adam優化器# 引入學習率調度器，在訓練過程中動態調整學習率--訓練初期使用較大的 LR 快速降低損失，訓練后期使用較小的 LR 更精細地逼近全局最優解。
# 在每個 epoch 結束后，需要手動調用調度器來更新學習率，可以在訓練過程中調用 scheduler.step()
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,        # 指定要控制的優化器（這里是Adam）mode='min',       # 監測的指標是"最小化"（如損失函數）patience=3,       # 如果連續3個epoch指標沒有改善，才降低LRfactor=0.5        # 降低LR的比例（新LR = 舊LR × 0.5）
)# 5. 訓練模型（記錄每個 iteration 的損失）
def train(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs):model.train()  # 設置為訓練模式# 記錄每個 iteration 的損失all_iter_losses = []  # 存儲所有 batch 的損失iter_indices = []     # 存儲 iteration 序號# 記錄每個 epoch 的準確率和損失train_acc_history = []test_acc_history = []train_loss_history = []test_loss_history = []for epoch in range(epochs):running_loss = 0.0correct = 0total = 0for batch_idx, (data, target) in enumerate(train_loader):data, target = data.to(device), target.to(device)  # 移至GPUoptimizer.zero_grad()  # 梯度清零output = model(data)  # 前向傳播loss = criterion(output, target)  # 計算損失loss.backward()  # 反向傳播optimizer.step()  # 更新參數# 記錄當前 iteration 的損失iter_loss = loss.item()all_iter_losses.append(iter_loss)iter_indices.append(epoch * len(train_loader) + batch_idx + 1)# 統計準確率和損失running_loss += iter_loss_, predicted = output.max(1)total += target.size(0)correct += predicted.eq(target).sum().item()# 每100個批次打印一次訓練信息if (batch_idx + 1) % 100 == 0:print(f'Epoch: {epoch+1}/{epochs} | Batch: {batch_idx+1}/{len(train_loader)} 'f'| 單Batch損失: {iter_loss:.4f} | 累計平均損失: {running_loss/(batch_idx+1):.4f}')# 計算當前epoch的平均訓練損失和準確率epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100. * correct / totaltrain_acc_history.append(epoch_train_acc)train_loss_history.append(epoch_train_loss)# 測試階段model.eval()  # 設置為評估模式test_loss = 0correct_test = 0total_test = 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += criterion(output, target).item()_, predicted = output.max(1)total_test += target.size(0)correct_test += predicted.eq(target).sum().item()epoch_test_loss = test_loss / len(test_loader)epoch_test_acc = 100. * correct_test / total_testtest_acc_history.append(epoch_test_acc)test_loss_history.append(epoch_test_loss)# 更新學習率調度器scheduler.step(epoch_test_loss)print(f'Epoch {epoch+1}/{epochs} 完成 | 訓練準確率: {epoch_train_acc:.2f}% | 測試準確率: {epoch_test_acc:.2f}%')# 繪制所有 iteration 的損失曲線plot_iter_losses(all_iter_losses, iter_indices)# 繪制每個 epoch 的準確率和損失曲線plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)return epoch_test_acc  # 返回最終測試準確率# 6. 繪制每個 iteration 的損失曲線
def plot_iter_losses(losses, indices):plt.figure(figsize=(10, 4))plt.plot(indices, losses, 'b-', alpha=0.7, label='Iteration Loss')plt.xlabel('Iteration（Batch序號）')plt.ylabel('損失值')plt.title('每個 Iteration 的訓練損失')plt.legend()plt.grid(True)plt.tight_layout()plt.show()# 7. 繪制每個 epoch 的準確率和損失曲線
def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):epochs = range(1, len(train_acc) + 1)plt.figure(figsize=(12, 4))# 繪制準確率曲線plt.subplot(1, 2, 1)plt.plot(epochs, train_acc, 'b-', label='訓練準確率')plt.plot(epochs, test_acc, 'r-', label='測試準確率')plt.xlabel('Epoch')plt.ylabel('準確率 (%)')plt.title('訓練和測試準確率')plt.legend()plt.grid(True)# 繪制損失曲線plt.subplot(1, 2, 2)plt.plot(epochs, train_loss, 'b-', label='訓練損失')plt.plot(epochs, test_loss, 'r-', label='測試損失')plt.xlabel('Epoch')plt.ylabel('損失值')plt.title('訓練和測試損失')plt.legend()plt.grid(True)plt.tight_layout()plt.show()# 訓練模型（復用原有的train函數）
print("開始訓練帶通道注意力的CNN模型...")
final_accuracy = train(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs=20)
print(f"訓練完成！最終測試準確率: {final_accuracy:.2f}%")

可視化空間注意力熱力圖

對比conv1,conv2,conv3這三個卷積層

# 可視化空間注意力熱力圖（顯示模型關注的圖像區域）
def visualize_attention_map(model, test_loader, device, class_names, num_samples=3):"""可視化模型的注意力熱力圖，展示模型關注的圖像區域"""model.eval()  # 設置為評估模式with torch.no_grad():for i, (images, labels) in enumerate(test_loader):if i >= num_samples:  # 只可視化前幾個樣本breakimages, labels = images.to(device), labels.to(device)# 為多個卷積層創建鉤子activation_maps = {}conv_layers = ['conv1', 'conv2', 'conv3']def hook(module, input, output, layer_name):activation_maps[layer_name] = output.cpu()# 為每個卷積層注冊鉤子hook_handles = []for layer_name in conv_layers:layer = getattr(model, layer_name)handle = layer.register_forward_hook(lambda m, i, o, name=layer_name: hook(m, i, o, name))hook_handles.append(handle)# 前向傳播，觸發鉤子outputs = model(images)# 移除所有鉤子for handle in hook_handles:handle.remove()# 獲取預測結果_, predicted = torch.max(outputs, 1)# 獲取原始圖像img = images[0].cpu().permute(1, 2, 0).numpy()# 反標準化處理img = img * np.array([0.2023, 0.1994, 0.2010]).reshape(1, 1, 3) + np.array([0.4914, 0.4822, 0.4465]).reshape(1, 1, 3)img = np.clip(img, 0, 1)# 為每個卷積層創建子圖for layer_name in conv_layers:# 獲取激活圖（對應卷積層的輸出）feature_map = activation_maps[layer_name][0].cpu()  # 取第一個樣本# 計算通道注意力權重（使用SE模塊的全局平均池化）channel_weights = torch.mean(feature_map, dim=(1, 2))  # [C]# 按權重對通道排序sorted_indices = torch.argsort(channel_weights, descending=True)# 創建子圖fig, axes = plt.subplots(1, 4, figsize=(16, 4))# 顯示原始圖像axes[0].imshow(img)axes[0].set_title(f'原始圖像\n真實: {class_names[labels[0]]}\n預測: {class_names[predicted[0]]}')axes[0].axis('off')# 顯示前3個最活躍通道的熱力圖for j in range(3):channel_idx = sorted_indices[j]# 獲取對應通道的特征圖channel_map = feature_map[channel_idx].numpy()# 歸一化到[0,1]channel_map = (channel_map - channel_map.min()) / (channel_map.max() - channel_map.min() + 1e-8)# 調整熱力圖大小以匹配原始圖像from scipy.ndimage import zoomheatmap = zoom(channel_map, (32/feature_map.shape[1], 32/feature_map.shape[2]))# 顯示熱力圖axes[j+1].imshow(img)axes[j+1].imshow(heatmap, alpha=0.5, cmap='jet')axes[j+1].set_title(f'{layer_name} 注意力熱力圖 - 通道 {channel_idx}')axes[j+1].axis('off')plt.tight_layout()plt.show()# 調用可視化函數
class_names = ['飛機', '汽車', '鳥', '貓', '鹿', '狗', '青蛙', '馬', '船', '卡車']
visualize_attention_map(model, test_loader, device, class_names, num_samples=3)

@浙大疏錦行