类别预测正确但特征图异常的问题求助

问题背景 #

基于ResNet50实现特征图可视化，模型类别预测结果正确，但生成的特征图存在异常：无论输入哪张图像，特征图都会在图像中部产生最强响应，且存在多区域异常响应。

实现代码 #

from torchvision.models.feature_extraction import create_feature_extractor

# Data processing
preprocess = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )])

image_path = './data/test_images/anemone.jpg'
image = Image.open(image_path).convert('RGB')
img_processed = preprocess(image)
batch_img_cat_tensor = torch.unsqueeze(img_processed, 0)

# Model initialization
resnet50_model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
# Eval model for predictions
resnet50_model.eval()

# Creating feature extractor
feature_extractor = create_feature_extractor(resnet50_model, 
                                             return_nodes=['layer4.2.conv3', 'fc'])

# Forward pass
out = feature_extractor(batch_img_cat_tensor)
pred = torch.argmax(out['fc'])

# Transforming last conv output to numpy and reshaping it so that the channels would be last
last_conv_output = torch.squeeze(out['layer4.2.conv3'])
last_conv_output = torch.reshape(last_conv_output, (7, 7, -1))
last_conv_output = last_conv_output.detach().numpy()
last_conv_output = last_conv_output.astype(np.uint8)

# Calculating the upscale factors for last conv output
width_factor = int(image.size[0] / last_conv_output.shape[0])
height_factor = int(image.size[1] / last_conv_output.shape[1])


# Getting the shapes of the last conv output
last_conv_w, last_conv_h, n_channels = last_conv_output.shape
# Calculate the 
upscaled_h = last_conv_h * height_factor
upscaled_w = last_conv_w * width_factor

# Upscaling the last_conv_output so that it could be "masked" with original image
upsampled_last_conv_output = np.zeros((upscaled_h, upscaled_w, n_channels))
upsampled_last_conv_output = []

for x in range(0, n_channels, 512):
    upsampled_last_conv_output.append(cv2.resize(last_conv_output[:, :, x:x+512], (upscaled_w, upscaled_h), cv2.INTER_CUBIC))
    
upsampled_last_conv_output = np.concatenate(upsampled_last_conv_output, axis=2)

# Getting the weights of the predicted class
last_layer_weights = resnet50_model.fc.weight.T
last_layer_weights_for_pred = last_layer_weights[:, pred]

# Dot multiplying the upsampled_last_conv_output with last_layer_weights_for_pred
upsampled_last_conv_output = upsampled_last_conv_output.reshape((-1, 2048))
heat_map = np.dot(upsampled_last_conv_output,
                  last_layer_weights_for_pred.detach().numpy()).reshape(upscaled_h, upscaled_w)

# Plotting the results
fig, ax = plt.subplots()
ax.imshow(image)
ax.imshow(heat_map, cmap='jet', alpha=0.5)
ax.set_title(prediction)

异常原因排查方向 #

• 特征值截断错误：代码中将卷积层输出直接转换为np.uint8，但ResNet卷积层ReLU后的输出数值通常远小于255，强制转换会导致大量特征值被截断为0，丢失有效信息，进而导致热力图响应异常。
• 热力图未归一化：生成的heat_map未做0-1或0-255范围的归一化处理，直接叠加到原图上时，数值范围异常会导致响应区域显示偏差。
• 冗余代码干扰：先初始化upsampled_last_conv_output为数组后又重新赋值为空列表，虽不影响运行，但可能增加调试复杂度，建议简化逻辑。

平台适配建议 #

该问题属于深度学习模型可视化的代码调试类问题，涉及PyTorch特征提取、热力图生成的技术细节，完全适合发布到AI Stack Exchange。发布时需附上完整代码、异常效果截图、已确认的正确预测结果，以便其他开发者快速定位问题。

内容的提问来源于stack exchange，提问作者artas2357