mmpretrain/docs/en/useful_tools/cam_visualization.md

Class Activation Map(CAM) Visualization

• Class Activation Map Visualization
  • Introduction of the CAM visualization tool
  • How to visualize the CAM of CNN(ResNet-50)
  • How to visualize the CAM of vision transformer

Introduction of the CAM visualization tool

MMClassification provides tools\visualizations\vis_cam.py tool to visualize class activation map. Please use pip install "grad-cam>=1.3.6" command to install pytorch-grad-cam.

The supported methods are as follows:

Method	What it does
GradCAM	Weight the 2D activations by the average gradient
GradCAM++	Like GradCAM but uses second order gradients
XGradCAM	Like GradCAM but scale the gradients by the normalized activations
EigenCAM	Takes the first principle component of the 2D Activations (no class discrimination, but seems to give great results)
EigenGradCAM	Like EigenCAM but with class discrimination: First principle component of Activations*Grad. Looks like GradCAM, but cleaner
LayerCAM	Spatially weight the activations by positive gradients. Works better especially in lower layers

Command：

python tools/visualizations/vis_cam.py \
    ${IMG} \
    ${CONFIG_FILE} \
    ${CHECKPOINT} \
    [--target-layers ${TARGET-LAYERS}] \
    [--preview-model] \
    [--method ${METHOD}] \
    [--target-category ${TARGET-CATEGORY}] \
    [--save-path ${SAVE_PATH}] \
    [--vit-like] \
    [--num-extra-tokens ${NUM-EXTRA-TOKENS}]
    [--aug_smooth] \
    [--eigen_smooth] \
    [--device ${DEVICE}] \
    [--cfg-options ${CFG-OPTIONS}]

Description of all arguments：

• img : The target picture path.
• config : The path of the model config file.
• checkpoint : The path of the checkpoint.
• --target-layers : The target layers to get activation maps, one or more network layers can be specified. If not set, use the norm layer of the last block.
• --preview-model : Whether to print all network layer names in the model.
• --method : Visualization method, supports GradCAM, GradCAM++, XGradCAM, EigenCAM, EigenGradCAM, LayerCAM, which is case insensitive. Defaults to GradCAM.
• --target-category : Target category, if not set, use the category detected by the given model.
• --save-path : The path to save the CAM visualization image. If not set, the CAM image will not be saved.
• --vit-like : Whether the network is ViT-like network.
• --num-extra-tokens : The number of extra tokens in ViT-like backbones. If not set, use num_extra_tokens the backbone.
• --aug_smooth : Whether to use TTA(Test Time Augment) to get CAM.
• --eigen_smooth : Whether to use the principal component to reduce noise.
• --device : The computing device used. Default to 'cpu'.
• --cfg-options : Modifications to the configuration file, refer to Learn about Configs.

The argument `--preview-model` can view all network layers names in the given model. It will be helpful if you know nothing about the model layers when setting `--target-layers`.

How to visualize the CAM of CNN(ResNet-50)

Here are some examples of target-layers in ResNet-50, which can be any module or layer:

• 'backbone.layer4' means the output of the forth ResLayer.
• 'backbone.layer4.2' means the output of the third BottleNeck block in the forth ResLayer.
• 'backbone.layer4.2.conv1' means the output of the conv1 layer in above BottleNeck block.

For `ModuleList` or `Sequential`, you can also use the index to specify which sub-module is the target layer.

For example, the `backbone.layer4[-1]` is the same as `backbone.layer4.2` since `layer4` is a `Sequential` with three sub-modules.

1. Use different methods to visualize CAM for ResNet50, the target-category is the predicted result by the given checkpoint, using the default target-layers.

   python tools/visualizations/vis_cam.py \
       demo/bird.JPEG \
       configs/resnet/resnet50_8xb32_in1k.py \
       https://download.openmmlab.com/mmclassification/v0/resnet/resnet50_batch256_imagenet_20200708-cfb998bf.pth \
       --method GradCAM
       # GradCAM++, XGradCAM, EigenCAM, EigenGradCAM, LayerCAM
   

   Image	GradCAM	GradCAM++	EigenGradCAM	LayerCAM

2. Use different target-category to get CAM from the same picture. In ImageNet dataset, the category 238 is 'Greater Swiss Mountain dog', the category 281 is 'tabby, tabby cat'.

   python tools/visualizations/vis_cam.py \
       demo/cat-dog.png configs/resnet/resnet50_8xb32_in1k.py \
       https://download.openmmlab.com/mmclassification/v0/resnet/resnet50_batch256_imagenet_20200708-cfb998bf.pth \
       --target-layers 'backbone.layer4.2' \
       --method GradCAM \
       --target-category 238
       # --target-category 281
   

   Category	Image	GradCAM	XGradCAM	LayerCAM
   Dog
   Cat

3. Use --eigen-smooth and --aug-smooth to improve visual effects.

   python tools/visualizations/vis_cam.py \
       demo/dog.jpg  \
       configs/mobilenet_v3/mobilenet-v3-large_8xb128_in1k.py \
       https://download.openmmlab.com/mmclassification/v0/mobilenet_v3/convert/mobilenet_v3_large-3ea3c186.pth \
       --target-layers 'backbone.layer16' \
       --method LayerCAM \
       --eigen-smooth --aug-smooth
   

   Image	LayerCAM	eigen-smooth	aug-smooth	eigen&aug

How to visualize the CAM of vision transformer

Here are some examples:

• 'backbone.norm3' for Swin-Transformer;
• 'backbone.layers[-1].ln1' for ViT;

For ViT-like networks, such as ViT, T2T-ViT and Swin-Transformer, the features are flattened. And for drawing the CAM, we need to specify the --vit-like argument to reshape the features into square feature maps.

Besides the flattened features, some ViT-like networks also add extra tokens like the class token in ViT and T2T-ViT, and the distillation token in DeiT. In these networks, the final classification is done on the tokens computed in the last attention block, and therefore, the classification score will not be affected by other features and the gradient of the classification score with respect to them, will be zero. Therefore, you shouldn't use the output of the last attention block as the target layer in these networks.

To exclude these extra tokens, we need know the number of extra tokens. Almost all transformer-based backbones in MMClassification have the num_extra_tokens attribute. If you want to use this tool in a new or third-party network that don't have the num_extra_tokens attribute, please specify it the --num-extra-tokens argument.

1. Visualize CAM for Swin Transformer, using default target-layers:

   python tools/visualizations/vis_cam.py \
       demo/bird.JPEG  \
       configs/swin_transformer/swin-tiny_16xb64_in1k.py \
       https://download.openmmlab.com/mmclassification/v0/swin-transformer/swin_tiny_224_b16x64_300e_imagenet_20210616_090925-66df6be6.pth \
       --vit-like
   

2. Visualize CAM for Vision Transformer(ViT):

   python tools/visualizations/vis_cam.py \
       demo/bird.JPEG  \
       configs/vision_transformer/vit-base-p16_ft-64xb64_in1k-384.py \
       https://download.openmmlab.com/mmclassification/v0/vit/finetune/vit-base-p16_in21k-pre-3rdparty_ft-64xb64_in1k-384_20210928-98e8652b.pth \
       --vit-like \
       --target-layers 'backbone.layers[-1].ln1'
   

3. Visualize CAM for T2T-ViT:

   python tools/visualizations/vis_cam.py \
       demo/bird.JPEG  \
       configs/t2t_vit/t2t-vit-t-14_8xb64_in1k.py \
       https://download.openmmlab.com/mmclassification/v0/t2t-vit/t2t-vit-t-14_3rdparty_8xb64_in1k_20210928-b7c09b62.pth \
       --vit-like \
       --target-layers 'backbone.encoder[-1].ln1'
   

Image	ResNet50	ViT	Swin	T2T-ViT