使用GradCAM对vision transformer进行注意力可视化的过程中遇到的问题记载

0.pip install grad-cam

1.首先是跑了其他一些博主的帖子，都没跑通，最后在知乎上找到的这篇帖子，修改了图片的输入设置后跑通了，实例代码如下

bash 复制代码

import cv2
import numpy as np
import torch

from pytorch_grad_cam import GradCAM, \
                            ScoreCAM, \
                            GradCAMPlusPlus, \
                            AblationCAM, \
                            XGradCAM, \
                            EigenCAM, \
                            EigenGradCAM, \
                            LayerCAM, \
                            FullGrad

from pytorch_grad_cam import GuidedBackpropReLUModel
from pytorch_grad_cam.utils.image import show_cam_on_image, preprocess_image

# 加载预训练的 ViT 模型
model = torch.hub.load('facebookresearch/deit:main','deit_tiny_patch16_224', pretrained=True)
model.eval()

# 判断是否使用 GPU 加速
use_cuda = torch.cuda.is_available()
if use_cuda:
    model = model.cuda()

def reshape_transform(tensor, height=14, width=14):
    # 去掉cls token
    result = tensor[:, 1:, :].reshape(tensor.size(0),
    height, width, tensor.size(2))

    # 将通道维度放到第一个位置
    result = result.transpose(2, 3).transpose(1, 2)
    return result

# 创建 GradCAM 对象
cam = GradCAM(model=model,
            target_layers=[model.blocks[-1].norm1],
            # 这里的target_layer要看模型情况，
            # 比如还有可能是：target_layers = [model.blocks[-1].ffn.norm]
            use_cuda=use_cuda,
            reshape_transform=reshape_transform)

# 读取输入图像
image_path = "2.png"
rgb_img = cv2.imread(image_path, 1)[:, :, ::-1]
rgb_img = cv2.resize(rgb_img, (224, 224))
rgb_img = np.float32(rgb_img) / 255.0

# 预处理图像
input_tensor = preprocess_image(rgb_img,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])

# 看情况将图像转换为批量形式
# input_tensor = input_tensor.unsqueeze(0)
if use_cuda:
    input_tensor = input_tensor.cuda()

# 计算 grad-cam
target_category = None # 可以指定一个类别，或者使用 None 表示最高概率的类别
grayscale_cam = cam(input_tensor=input_tensor, targets=target_category)
grayscale_cam = grayscale_cam[0, :]

# 将 grad-cam 的输出叠加到原始图像上
visualization = show_cam_on_image(rgb_img, grayscale_cam)

# 保存可视化结果
cv2.cvtColor(visualization, cv2.COLOR_RGB2BGR, visualization)
cv2.imwrite('cam3.jpg', visualization)

参考链接：

bash 复制代码

https://zhuanlan.zhihu.com/p/640450435

2.接下来参考了这位大佬的博客，针对我的网络进行修改，一些遇到的错误在大佬在里面说了

bash 复制代码

https://blog.csdn.net/holly_Z_P_F/article/details/130011296

3.新的问题

AttributeError: 'NoneType' object has no attribute 'shape'

bash 复制代码

Traceback (most recent call last):
  File "D:\yanjiusheng\ZSE-SBIR\kk.py", line 68, in <module>
    test()
  File "D:\yanjiusheng\ZSE-SBIR\kk.py", line 56, in test
    grayscale_cam = cam(input_tensor=input_tensor, targets=target_category)
  File "D:\anaconda3\envs\zse-sbir\lib\site-packages\pytorch_grad_cam\base_cam.py", line 186, in __call__
    return self.forward(input_tensor, targets, eigen_smooth)
  File "D:\anaconda3\envs\zse-sbir\lib\site-packages\pytorch_grad_cam\base_cam.py", line 110, in forward
    cam_per_layer = self.compute_cam_per_layer(input_tensor, targets, eigen_smooth)
  File "D:\anaconda3\envs\zse-sbir\lib\site-packages\pytorch_grad_cam\base_cam.py", line 141, in compute_cam_per_layer
    cam = self.get_cam_image(input_tensor, target_layer, targets, layer_activations, layer_grads, eigen_smooth)
  File "D:\anaconda3\envs\zse-sbir\lib\site-packages\pytorch_grad_cam\base_cam.py", line 66, in get_cam_image
    weights = self.get_cam_weights(input_tensor, target_layer, targets, activations, grads)
  File "D:\anaconda3\envs\zse-sbir\lib\site-packages\pytorch_grad_cam\grad_cam.py", line 23, in get_cam_weights
    if len(grads.shape) == 4:
AttributeError: 'NoneType' object has no attribute 'shape'

这个问题我最后解决也是一知半解做出来了，目标层的问题:GradCAM 需要一个可以计算梯度的层作为 target_layer，通常是卷积层或自注意力机制中的特定部分，我原来选择的是model.sa.model.transformer.layers[-1] 作为目标层，但是这可能并不是一个适合用于 GradCAM 的层。我选择了原来目标层下更具体的一层，即model.sa.model.transformer.layers[-1][0].layer_norm_input ，兄弟们可以 print（model） 看下结构，多试几层看看哪层能用，我太菜了只能试。。。

参考的另外一些帖子：

bash 复制代码

https://ask.csdn.net/questions/8051598

bash 复制代码

https://github.com/open-mmlab/mmdetection/issues/1809