yidingelaine / fullgrad-saliency

Full-gradient saliency maps

Full-Gradient Saliency Maps

This is the reference implementation of the FullGrad saliency method described in our NeurIPS 2019 publication "Full-Gradient Representation for Neural Network Visualization".

This repository implements three methods:

1. The reference FullGrad saliency method, which aggregates layerwise gradient maps multipled with the bias terms
2. Simple FullGrad, which omits computation of bias parameters and simply aggregates layerwise gradient maps
3. Smooth FullGrad, which aggregates noise averaged layerwise gradient maps with the bias terms

The related full-gradient decomposition is implemented within FullGrad. Note that while full-gradient decomposition applies to any ReLU neural network, FullGrad saliency is specific to CNNs.

In addition, the following methods from literature are also implemented:

1. Input-gradient (paper)
2. Grad-CAM (paper)
3. SmoothGrad (paper)

Examples

FullGrad	Simple FullGrad	Smooth FullGrad

Grad-CAM	Input-Gradients	SmoothGrad

Usage

Simply run python dump_images.py, the saliency maps should be saved consequently in a results folder.

Interfaces

The FullGrad class has the following methods implemented.

from saliency.fullGrad import FullGrad

# Initialize FullGrad object
fullgrad = FullGrad(model)

# Check completeness property
# done automatically while initializing object
fullgrad.checkCompleteness()

# Obtain fullgradient decomposition
input_gradient, bias_gradients = 
fullgrad.fullGradientDecompose(input_image, target_class)

# Obtain saliency maps
saliency_map = fullgrad.saliency(input_image, target_class)

We also introduce variants called SimpleFullGrad and SmoothFullGrad which have no completeness property or decomposition.

from saliency.simple_fullgrad import SimpleFullGrad
from saliency.smooth_fullgrad import SmoothFullGrad

# Initialize Simple / Smooth FullGrad objects
simple_fullgrad = SimpleFullGrad(model)
smooth_fullgrad = SmoothFullGrad(model)

# Obtain saliency maps
saliency_map_simple = simple_fullgrad.saliency(input_image, target_class)
saliency_map_smooth = smooth_fullgrad.saliency(input_image, target_class)

This basic interface is retained for input-gradient, gradcam and smoothgrad methods as well.

Dependencies

torch torchvision cv2 numpy 

Research

If you found our work helpful for your research, please do consider citing us.

@inproceedings{srinivas2019fullgrad,
    title={Full-Gradient Representation for Neural Network Visualization},
    author={Srinivas, Suraj and Fleuret, François},
    booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
    year={2019}
}