Python打卡训练营学习记录Day46
作业：
- 今日代码较多，理解逻辑即可
- 对比不同卷积层特征图可视化的结果（可选）
一、CNN特征图可视化实现
import torch
import matplotlib.pyplot as pltdef visualize_feature_maps(model, input_tensor):# 注册钩子获取中间层输出features = []def hook(module, input, output):features.append(output.detach().cpu())# 选择不同卷积层观察target_layers = [model.layer1[0].conv1,model.layer2[0].conv1,model.layer3[0].conv1]handles = []for layer in target_layers:handles.append(layer.register_forward_hook(hook))# 前向传播with torch.no_grad():_ = model(input_tensor.unsqueeze(0))# 移除钩子for handle in handles:handle.remove()# 可视化不同层特征图fig, axes = plt.subplots(len(target_layers), 5, figsize=(20, 10))for i, feat in enumerate(features):for j in range(5): # 显示前5个通道axes[i,j].imshow(feat[0, j].numpy(), cmap='viridis')axes[i,j].axis('off')plt.show()二、通道注意力模块示例
class ChannelAttention(nn.Module):def __init__(self, in_channels, reduction=16):super().__init__()self.avg_pool = nn.AdaptiveAvgPool2d(1)self.max_pool = nn.AdaptiveMaxPool2d(1)self.fc = nn.Sequential(nn.Linear(in_channels, in_channels // reduction),nn.ReLU(),nn.Linear(in_channels // reduction, in_channels),nn.Sigmoid())def forward(self, x):# ... existing code ...return x * attention_weights # 应用注意力权重三、热力图生成方法
def generate_heatmap(model, input_img):# 前向传播获取梯度model.eval()input_img.requires_grad = Trueoutput = model(input_img)pred_class = output.argmax(dim=1).item()# 反向传播计算梯度model.zero_grad()output[0, pred_class].backward()# 获取最后一个卷积层的梯度gradients = model.layer4[1].conv2.weight.gradpooled_gradients = torch.mean(gradients, dim=[0,2,3])# 生成热力图activations = model.layer4[1].conv2.activations.detach()for i in range(activations.shape[1]):activations[:,i,:,:] *= pooled_gradients[i]heatmap = torch.mean(activations, dim=1).squeeze()return heatmap@浙大疏锦行