思路

获取想要解释的那一层的特征图，然后根据特征图梯度计算出权重值，加在原图上面。

Demo

加上类激活(cam)

可以看到，cam将模型认为有利于分类的特征标注了出来。
下面以ResNet50为例:
Trick:
使用

for i in model._modules.items():

可以获得模型名称和对应层。

# coding: utf-8
import os
import cv2
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
 
import torch
import torch.autograd as autograd
import torchvision.transforms as transforms

import torchvision.models as models
 
 
# 训练过的模型路径
#resume_path = r"D:\TJU\GBDB\set113\cross_validation\test1\epoch_0257_checkpoint.pth.tar"
# 输入图像路径
single_img_path = r'bicycle.jpg'
# 绘制的热力图存储路径
save_path = r'heatmap/bicycle_layer4.jpg'
 
# 网络层的层名列表, 需要根据实际使用网络进行修改
layers_names = ['conv1', 'bn1', 'relu', 'maxpool', 'layer1', 'layer2', 'layer3', 'layer4', 'avgpool']
# 指定层名
out_layer_name = "layer4"
 
features_grad = 0
 
 
# 为了读取模型中间参数变量的梯度而定义的辅助函数
def extract(g):
    global features_grad
    features_grad = g
 
 
def draw_CAM(model, img_path, save_path, transform=None, visual_heatmap=False, out_layer=None):
    """
    绘制 Class Activation Map
    :param model: 加载好权重的Pytorch model
    :param img_path: 测试图片路径
    :param save_path: CAM结果保存路径
    :param transform: 输入图像预处理方法
    :param visual_heatmap: 是否可视化原始heatmap（调用matplotlib）
    :return:
    """
    # 读取图像并预处理
    global layer2
    img = Image.open(img_path).convert('RGB')
    if transform:
        img = transform(img)
    img = img.unsqueeze(0)  # (1, 3, 448, 448)
 
    # model转为eval模式
    model.eval()
 
    # 获取模型层的字典
    layers_dict = {layers_names[i]: None for i in range(len(layers_names))}
    for name,module in model._modules.items():
        #print(i, (name, module))
        layers_dict[name] = module
 
    # 遍历模型的每一层, 获得指定层的输出特征图
    # features: 指定层输出的特征图, features_flatten: 为继续完成前端传播而设置的变量
    features = img
    start_flatten = False
    features_flatten = None

    for name, layer in layers_dict.items():
        if name != out_layer and start_flatten is False:    # 指定层之前
            features = layer(features)
        elif name == out_layer and start_flatten is False:  # 指定层
            features = layer(features)
            start_flatten = True
        else:   # 指定层之后
            if name == "fc":
                break
            if features_flatten is None:
                features_flatten = layer(features)
            else:
                features_flatten = layer(features_flatten)
    #print(features_flatten.shape)
    features_flatten = torch.flatten(features_flatten, 1)
    #print(features_flatten.shape)
    output = model.fc(features_flatten)
    # 预测得分最高的那一类对应的输出score
    pred = torch.argmax(output, 1).item()
    pred_class = output[:, pred]
 
    # 求中间变量features的梯度
    # 方法1
    # features.register_hook(extract)
    # pred_class.backward()
    # 方法2
    features_grad = autograd.grad(pred_class, features, allow_unused=True)[0]
 
    grads = features_grad  # 获取梯度
    pooled_grads = torch.nn.functional.adaptive_avg_pool2d(grads, (1, 1))
    # 此处batch size默认为1，所以去掉了第0维（batch size维）
    pooled_grads = pooled_grads[0]
    features = features[0]
    print("pooled_grads:", pooled_grads.shape)
    print("features:", features.shape)
    # features.shape[0]是指定层feature的通道数
    for i in range(features.shape[0]):
        features[i, ...] *= pooled_grads[i, ...]
 
    # 计算heatmap
    heatmap = features.detach().cpu().numpy()
    heatmap = np.mean(heatmap, axis=0)
    heatmap = np.maximum(heatmap, 0)
    heatmap /= np.max(heatmap)
 
    # 可视化原始热力图
    if visual_heatmap:
        plt.matshow(heatmap)
        plt.show()
 
    img = cv2.imread(img_path)  # 用cv2加载原始图像
    heatmap = cv2.resize(heatmap, (img.shape[1], img.shape[0]))  # 将热力图的大小调整为与原始图像相同
    heatmap = np.uint8(255 * heatmap)  # 将热力图转换为RGB格式
    heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)  # 将热力图应用于原始图像
    superimposed_img = heatmap * 0.7 + img  # 这里的0.4是热力图强度因子
    cv2.imwrite(save_path, superimposed_img)  # 将图像保存到硬盘
 
 
if __name__ == '__main__':
    model = models.resnet50(pretrained=True)
    #model.eval()
    transform = transforms.Compose([
        transforms.Resize(448),
        transforms.ToTensor(),
        transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
    ])
    # 构建模型并加载预训练参数
    #seresnet50 = FineTuneSEResnet50(num_class=113).cuda()
    #checkpoint = torch.load(resume_path)
    #seresnet50.load_state_dict(checkpoint['state_dict'])
    draw_CAM(model, single_img_path, save_path, transform=transform, visual_heatmap=True, out_layer=out_layer_name)