Advanced AI explainability for PyTorch

pip install grad-cam

Documentation with advanced tutorials: https://jacobgil.github.io/pytorch-gradcam-book

This is a package with state of the art methods for Explainable AI for computer vision. This can be used for diagnosing model predictions, either in production or while developing models. The aim is also to serve as a benchmark of algorithms and metrics for research of new explainability methods.

⭐ Comprehensive collection of Pixel Attribution methods for Computer Vision.

⭐ Tested on many Common CNN Networks and Vision Transformers.

⭐ Advanced use cases: Works with Classification, Object Detection, Semantic Segmentation, Embedding-similarity and more.

⭐ Includes smoothing methods to make the CAMs look nice.

⭐ High performance: full support for batches of images in all methods.

⭐ Includes metrics for checking if you can trust the explanations, and tuning them for best performance.

Method	What it does
GradCAM	Weight the 2D activations by the average gradient
HiResCAM	Like GradCAM but element-wise multiply the activations with the gradients; provably guaranteed faithfulness for certain models
GradCAMElementWise	Like GradCAM but element-wise multiply the activations with the gradients then apply a ReLU operation before summing
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
LayerCAM	Spatially weight the activations by positive gradients. Works better especially in lower layers
FullGrad	Computes the gradients of the biases from all over the network, and then sums them
Deep Feature Factorizations	Non Negative Matrix Factorization on the 2D activations

Visual Examples

What makes the network think the image label is 'pug, pug-dog'	What makes the network think the image label is 'tabby, tabby cat'	Combining Grad-CAM with Guided Backpropagation for the 'pug, pug-dog' class

Object Detection and Semantic Segmentation

Object Detection	Semantic Segmentation

3D Medical Semantic Segmentation

Explaining similarity to other images / embeddings

Deep Feature Factorization

Classification

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

Metrics and Evaluation for XAI

---

Usage examples

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

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

# We have to specify the target we want to generate the CAM for.
targets = [ClassifierOutputTarget(281)]

# Construct the CAM object once, and then re-use it on many images.
with GradCAM(model=model, target_layers=target_layers) as cam:
  # You can also pass aug_smooth=True and eigen_smooth=True, to apply smoothing.
  grayscale_cam = cam(input_tensor=input_tensor, targets=targets)
  # 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, use_rgb=True)
  # You can also get the model outputs without having to redo inference
  model_outputs = cam.outputs

cam.py has a more detailed usage example.

---

Choosing the layer(s) to extract activations from

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

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

If you pass a list with several layers, the CAM will be averaged accross them. This can be useful if you're not sure what layer will perform best.

---

Adapting for new architectures and tasks

Methods like GradCAM were designed for and were originally mostly applied on classification models, and specifically CNN classification models. However you can also use this package on new architectures like Vision Transformers, and on non classification tasks like Object Detection or Semantic Segmentation.

The be able to adapt to non standard cases, we have two concepts.

• The reshape transform - how do we convert activations to represent spatial images ?
• The model targets - What exactly should the explainability method try to explain ?

The reshape_transform argument

In a CNN the intermediate activations in the model are a mult-channel image that have the dimensions channel x rows x cols, and the various explainabiltiy methods work with these to produce a new image.

In case of another architecture, like the Vision Transformer, the shape might be different, like (rows x cols + 1) x channels, or something else. The reshape transform converts the activations back into a multi-channel image, for example by removing the class token in a vision transformer. For examples, check here

The model_target argument

The model target is just a callable that is able to get the model output, and filter it out for the specific scalar output we want to explain.

For classification tasks, the model target will typically be the output from a specific category. The targets parameter passed to the CAM method can then use ClassifierOutputTarget:

targets = [ClassifierOutputTarget(281)]

However for more advanced cases, you might want a different behaviour. Check here for more examples.

---

Tutorials

Here you can find detailed examples of how to use this for various custom use cases like object detection:

These point to the new documentation jupter-book for fast rendering. The jupyter notebooks themselves can be found under the tutorials folder in the git repository.

• Notebook tutorial: XAI Recipes for the HuggingFace 🤗 Image Classification Models

• Notebook tutorial: Deep Feature Factorizations for better model explainability

• Notebook tutorial: Class Activation Maps for Object Detection with Faster-RCNN

• Notebook tutorial: Class Activation Maps for YOLO5

• Notebook tutorial: Class Activation Maps for Semantic Segmentation

• Notebook tutorial: Adapting pixel attribution methods for embedding outputs from models

• Notebook tutorial: May the best explanation win. CAM Metrics and Tuning

• How it works with Vision/SwinT transformers

---

Guided backpropagation

from pytorch_grad_cam import GuidedBackpropReLUModel
from pytorch_grad_cam.utils.image import (
    show_cam_on_image, deprocess_image, preprocess_image
)
gb_model = GuidedBackpropReLUModel(model=model, device=model.device())
gb = gb_model(input_tensor, target_category=None)

cam_mask = cv2.merge([grayscale_cam, grayscale_cam, grayscale_cam])
cam_gb = deprocess_image(cam_mask * gb)
result = deprocess_image(gb)

---

Metrics and evaluating the explanations

from pytorch_grad_cam.utils.model_targets import ClassifierOutputSoftmaxTarget
from pytorch_grad_cam.metrics.cam_mult_image import CamMultImageConfidenceChange
# Create the metric target, often the confidence drop in a score of some category
metric_target = ClassifierOutputSoftmaxTarget(281)
scores, batch_visualizations = CamMultImageConfidenceChange()(input_tensor, 
  inverse_cams, targets, model, return_visualization=True)
visualization = deprocess_image(batch_visualizations[0, :])

# State of the art metric: Remove and Debias
from pytorch_grad_cam.metrics.road import ROADMostRelevantFirst, ROADLeastRelevantFirst
cam_metric = ROADMostRelevantFirst(percentile=75)
scores, perturbation_visualizations = cam_metric(input_tensor, 
  grayscale_cams, targets, model, return_visualization=True)

# You can also average across different percentiles, and combine
# (LeastRelevantFirst - MostRelevantFirst) / 2
from pytorch_grad_cam.metrics.road import ROADMostRelevantFirstAverage,
                                          ROADLeastRelevantFirstAverage,
                                          ROADCombined
cam_metric = ROADCombined(percentiles=[20, 40, 60, 80])
scores = cam_metric(input_tensor, grayscale_cams, targets, model)

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> --output-dir <output_dir_path>

To use with a specific device, like cpu, cuda, cuda:0 or mps: python cam.py --image-path <path_to_image> --device cuda --output-dir <output_dir_path>

---

You can choose between:

GradCAM , HiResCAM, ScoreCAM, GradCAMPlusPlus, AblationCAM, XGradCAM , LayerCAM, FullGrad 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 =

---

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/2011.08891
Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks Rachel L. Draelos, Lawrence Carin

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

http://mftp.mmcheng.net/Papers/21TIP_LayerCAM.pdf
LayerCAM: Exploring Hierarchical Class Activation Maps for Localization Peng-Tao Jiang; Chang-Bin Zhang; Qibin Hou; Ming-Ming Cheng; Yunchao Wei

https://arxiv.org/abs/1905.00780
Full-Gradient Representation for Neural Network Visualization Suraj Srinivas, Francois Fleuret

https://arxiv.org/abs/1806.10206
Deep Feature Factorization For Concept Discovery Edo Collins, Radhakrishna Achanta, Sabine Süsstrunk