特征可视化之CAM

1 介绍

类别激活映射Class Activation Maps（CAM）可用于可视化深度学习模型中间层特征图，它通过高亮输入图像的部分区域（对于模型识别目标的贡献较大），提供了对模型决策效果的直观解释。论文链接：CAM

具体操作：

CAM由两部分加权构成：原图+特征图，其中，特征图是通过：最后一个全连接层的参数与最后输出的特征图对应相乘再相加而成，如下图所示：

2 基于ConvNeXt可视化代码

import torch
import cv2
import numpy as np
from PIL import Image
import torchvision.transforms as transforms


def draw_feature(model, img_path, save_img_path):
    origin_image = Image.open(img_path)
    input_shape = [224, 224]
    data_transform = transforms.Compose(
        [transforms.Resize(input_shape),
         transforms.ToTensor(),
         transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])

    image = data_transform(origin_image)
    x = torch.unsqueeze(image, dim=0)
    x = model.downsample_layers[0](x)
    x = model.stages[0](x)
    x = model.downsample_layers[1](x)
    x = model.stages[1](x)
    x = model.downsample_layers[2](x)
    x = model.stages[2](x)
    x = model.downsample_layers[3](x)
    out = model.stages[3](x)    # 1×768×7×7
    the_feature = out

    out = model.norm(out.mean([-2, -1]))
    out = model.head(out)
    def extract(g):
        global feature_grad
        feature_grad = g

    pred = torch.argmax(out).item()
    pred_class = out[:, pred]
    the_feature.register_hook(extract)
    pred_class.backward()
    greds = feature_grad
    pooled_grads = torch.nn.functional.adaptive_avg_pool2d(greds, (1, 1))
    pooled_grads = pooled_grads[0]
    print(pooled_grads.shape)
    features = the_feature[0].cpu()
    for i in range(the_feature.shape[1]):
        features[i, ...] *= pooled_grads[i, ...]
    print(features.shape)
    heatmap = features.detach().numpy()
    print(heatmap.shape)
    heatmap = np.mean(heatmap, axis=0)
    print(heatmap.shape)
    heatmap /= np.max(heatmap)

    img = cv2.imread(img_path)
    heatmap = cv2.resize(heatmap, (img.shape[1], img.shape[0]))
    heatmap = np.uint8(255 * heatmap)
    heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)
    superimposed_img = heatmap * 0.4 + img
    cv2.imwrite(save_img_path, superimposed_img)


if __name__ == '__main__':
    img_path = "images/daisy.jpg"
    model_path = "logs/best_model.pth"
    save_pic_path = "./cam.jpg"
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    model = torch.load(model_path, map_location='cpu')
    model.eval()
    draw_feature(model, img_path, save_pic_path)

可视化效果：