April 28, 2021

1105 words 6 mins read

jacobgil/pytorch-grad-cam

jacobgil/pytorch-grad-cam

Many Class Activation Map methods implemented in Pytorch for CNNs and Vision Transformers. Including Grad-CAM, Grad-CAM++, Score-CAM, Ablation-CAM and XGrad-CAM

repo name jacobgil/pytorch-grad-cam
repo link https://github.com/jacobgil/pytorch-grad-cam
homepage
language Python
size (curr.) 1895 kB
stars (curr.) 2434
created 2017-05-31
license MIT License

Class Activation Map methods implemented in Pytorch

pip install grad-cam

⭐ Tested on many Common CNN Networks and Vision Transformers.

⭐ Includes smoothing methods to make the CAMs look nice.

⭐ Full support for batches of images in all methods.

visualization

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
AblationCAM Zero out activations and measure how the output drops (this repository includes a fast batched implementation)
ScoreCAM Perbutate the image by the scaled activations and measure how the output drops
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

What makes the network think the image label is ‘pug, pug-dog’ and ‘tabby, tabby cat’:

Dog Cat

Combining Grad-CAM with Guided Backpropagation for the ‘pug, pug-dog’ class:

Combined

More Visual Examples

Resnet50:

Category Image GradCAM AblationCAM ScoreCAM
Dog
Cat

Vision Transfomer (Deit Tiny):

Category Image GradCAM AblationCAM ScoreCAM
Dog
Cat

Swin Transfomer (Tiny window:7 patch:4 input-size:224):

Category Image GradCAM AblationCAM ScoreCAM
Dog
Cat

It seems that GradCAM++ is almost the same as GradCAM, in most networks except VGG where the advantage is larger.

Network Image GradCAM GradCAM++ Score-CAM Ablation-CAM Eigen-CAM
VGG16
Resnet50

Chosing the Target Layer

You need to choose the target layer to compute CAM for. Some common choices are:

  • Resnet18 and 50: model.layer4[-1]
  • VGG and densenet161: model.features[-1]
  • mnasnet1_0: model.layers[-1]
  • ViT: model.blocks[-1].norm1
  • SwinT: model.layers[-1].blocks[-1].norm1

Using from code as a library

from pytorch_grad_cam import GradCAM, ScoreCAM, GradCAMPlusPlus, AblationCAM, XGradCAM, EigenCAM
from pytorch_grad_cam.utils.image import show_cam_on_image
from torchvision.models import resnet50

model = resnet50(pretrained=True)
target_layer = model.layer4[-1]
input_tensor = # Create an input tensor image for your model..
# Note: input_tensor can be a batch tensor with several images!

# Construct the CAM object once, and then re-use it on many images:
cam = GradCAM(model=model, target_layer=target_layer, use_cuda=args.use_cuda)

# If target_category is None, the highest scoring category
# will be used for every image in the batch.
# target_category can also be an integer, or a list of different integers
# for every image in the batch.
target_category = 281

# You can also pass aug_smooth=True and eigen_smooth=True, to apply smoothing.
grayscale_cam = cam(input_tensor=input_tensor, target_category=target_category)

# In this example grayscale_cam has only one image in the batch:
grayscale_cam = grayscale_cam[0, :]
visualization = show_cam_on_image(rgb_img, grayscale_cam)

Smoothing to get nice looking CAMs

To reduce noise in the CAMs, and make it fit better on the objects, two smoothing methods are supported:

  • aug_smooth=True

    Test time augmentation: increases the run time by x6.

    Applies a combination of horizontal flips, and mutiplying the image by [1.0, 1.1, 0.9].

    This has the effect of better centering the CAM around the objects.

  • eigen_smooth=True

    First principle component of activations*weights

    This has the effect of removing a lot of noise.

AblationCAM aug smooth eigen smooth aug+eigen smooth

Running the example script:

Usage: python cam.py --image-path <path_to_image> --method <method>

To use with CUDA: python cam.py --image-path <path_to_image> --use-cuda


You can choose between:

GradCAM , ScoreCAM, GradCAMPlusPlus, AblationCAM, XGradCAM and EigenCAM.

Some methods like ScoreCAM and AblationCAM require a large number of forward passes, and have a batched implementation.

You can control the batch size with cam.batch_size =


How does it work with Vision Transformers

See usage_examples/vit_example.py

In ViT the output of the layers are typically BATCH x 197 x 192. In the dimension with 197, the first element represents the class token, and the rest represent the 14x14 patches in the image. We can treat the last 196 elements as a 14x14 spatial image, with 192 channels.

To reshape the activations and gradients to 2D spatial images, we can pass the CAM constructor a reshape_transform function.

This can also be a starting point for other architectures that will come in the future.


GradCAM(model=model, target_layer=target_layer, reshape_transform=reshape_transform)

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

    # Bring the channels to the first dimension,
    # like in CNNs.
    result = result.transpose(2, 3).transpose(1, 2)
    return result

Which target_layer should we chose for Vision Transformers?

Since the final classification is done on the class token computed in the last attention block, the output will not be affected by the 14x14 channels in the last layer. The gradient of the output with respect to them, will be 0!

We should chose any layer before the final attention block, for example:

target_layer = model.blocks[-1].norm1

How does it work with Swin Transformers

See usage_examples/swinT_example.py

In Swin transformer base the output of the layers are typically BATCH x 49 x 1024. We can treat the last 49 elements as a 7x7 spatial image, with 1024 channels.

To reshape the activations and gradients to 2D spatial images, we can pass the CAM constructor a reshape_transform function.

This can also be a starting point for other architectures that will come in the future.


GradCAM(model=model, target_layer=target_layer, reshape_transform=reshape_transform)

def reshape_transform(tensor, height=7, width=7):
    result = tensor.reshape(tensor.size(0),
        height, width, tensor.size(2))

    # Bring the channels to the first dimension,
    # like in CNNs.
    result = result.transpose(2, 3).transpose(1, 2)
    return result

Which target_layer should we chose for Swin Transformers?

Since the swin transformer is different from ViT, it does not contains cls_token as present in ViT, therefore we will use all the 7x7 images we get from the last block of the last layer.

We should chose any layer before the final attention block, for example:

target_layer = model.layers[-1].blocks[-1].norm1

Citation

If you use this for research, please cite. Here is an example BibTeX entry:

@misc{jacobgilpytorchcam,
  title={PyTorch library for CAM methods},
  author={Jacob Gildenblat and contributors},
  year={2021},
  publisher={GitHub},
  howpublished={\url{https://github.com/jacobgil/pytorch-grad-cam}},
}

References

https://arxiv.org/abs/1610.02391 Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra

https://arxiv.org/abs/1710.11063 Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, Vineeth N Balasubramanian

https://arxiv.org/abs/1910.01279 Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, Xia Hu

https://ieeexplore.ieee.org/abstract/document/9093360/ Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. Saurabh Desai and Harish G Ramaswamy. In WACV, pages 972–980, 2020

https://arxiv.org/abs/2008.02312 Axiom-based Grad-CAM: Towards Accurate Visualization and Explanation of CNNs Ruigang Fu, Qingyong Hu, Xiaohu Dong, Yulan Guo, Yinghui Gao, Biao Li

https://arxiv.org/abs/2008.00299 Eigen-CAM: Class Activation Map using Principal Components Mohammed Bany Muhammad, Mohammed Yeasin

comments powered by Disqus