https://github.com/hummat/saliency

PyTorch implementation of 'Vanilla' Gradient, Grad-CAM, Guided backprop, Integrated Gradients and their SmoothGrad variants.
https://github.com/hummat/saliency

deep-learning deep-learning-algorithms deep-neural-networks grad-cam guided-backpropagation integrated-gradients machine-learning machine-learning-algorithms pytorch saliency smoothgrad xrai

Last synced: 1 day ago
JSON representation

PyTorch implementation of 'Vanilla' Gradient, Grad-CAM, Guided backprop, Integrated Gradients and their SmoothGrad variants.

• Host: GitHub
• URL: https://github.com/hummat/saliency
• Owner: hummat
• License: gpl-3.0
• Created: 2019-12-21T15:53:36.000Z (almost 6 years ago)
• Default Branch: master
• Last Pushed: 2024-12-09T13:59:48.000Z (10 months ago)
• Last Synced: 2025-06-06T16:11:38.168Z (4 months ago)
• Topics: deep-learning, deep-learning-algorithms, deep-neural-networks, grad-cam, guided-backpropagation, integrated-gradients, machine-learning, machine-learning-algorithms, pytorch, saliency, smoothgrad, xrai
• Language: Jupyter Notebook
• Homepage: https://hummat.github.io/saliency/
• Size: 9.83 MB
• Stars: 19
• Watchers: 0
• Forks: 3
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Saliency Methods

## Overview

This repository contains code for the following saliency techniques:

### 1. Vanilla Gradients

![gradient](images/gradient.png)

### 2. Guided Backpropogation (with Smoothgrad)

![guided](images/guided_backrop_smoothgrad.png)

### 3. Integrated Gradients

![integrated](images/integrated_gradient.png)

### 4. (Guided) Grad-CAM

![gradcam](images/guided_grad_cam.png)

### 5. XRAI

![xrai](images/xrai.png)

## Remarks

The methods should work with all models from the [torchvision](https://github.com/pytorch/vision) package. Tested models so far are:

* VGG variants

* ResNet variants

* DenseNet variants

* Inception/GoogLeNet*

*In order for *Guided Backpropagation* and *Grad-CAM* to work properly with the *Inception* and *GoogLeNet* models, they need to by modified slightly, such that all *ReLUs* are modules of the model rather than function calls.

```python

# This class can be found at the very end of inception.py and googlenet.py respectively.

class BasicConv2d(nn.Module):

    def __init__(self, in_channels, out_channels, **kwargs):

        super(BasicConv2d, self).__init__()

        self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs)

        self.bn = nn.BatchNorm2d(out_channels, eps=0.001)

        self.relu = nn.ReLU(inplace=True)  # Add this line

    def forward(self, x):

        x = self.conv(x)

        x = self.bn(x)

        return self.relu(x)  # Replaces F.relu(x, inplace=True)

```

    

## Examples

For a brief overview on how to use the package, please have a look at this short [tutorial notebook](https://github.com/hummat/saliency/blob/master/tutorial.ipynb). The bare minimum is summarized below.

```python

# Standard imports 

import torchvision

# Import desired utils and methods

from ml_utils import load_image, show_mask

from guided_backprop import GuidedBackprop

# Load model and image

model = torchvision.models.resnet50(pretrained=True)

doberman = load_image('images/doberman.png', size=224)

# Construct a saliency object and compute the saliency mask.

guided_backprop = GuidedBackprop(model)

rgb_mask = guided_backprop.get_mask(image_tensor=doberman)

# Visualize the result

show_mask(rgb_mask, title='Guided Backprop')

```

## Credits

The implementation follows closely that of the corresponding [TensorFlow saliency](https://github.com/PAIR-code/saliency) repository, reusing its code were applicable (mostly for the XRAI method).

Further inspiration has been taken from [this](https://github.com/utkuozbulak/pytorch-cnn-visualizations) repository.