Some simple codes for the paper https://arxiv.org/abs/1805.07039

Requirement.

We suggest you run the platform under Python 3.6+ with following libs:

• TensorFlow >= 1.4.0
• Numpy 1.12.1
• Scipy 0.19.0

Get Started

1. Visualize the fully-connected networks (FC):

cd fc
# default setting: Guided Backpropagation (GBP) with the max logit
python3 visualize_fc.py

2. Visualize the three-layer CNN:

cd convnet
# default setting: Guided Backpropagation (GBP) with the max logit
python3 visualize_convnet.py

3. Visualize the VGG-16 net:

cd deepnet
# default setting: Guided Backpropagation (GBP) with the max logit
python3 visualize_vgg.py

Note: In the visualize_[fc, convnet, vgg].py, you could change the variable "sal_type" from ['GuidedBackprop', 'Deconv', 'PlainSaliency'] to get different visualizations, and change the variable "logit_type" from ['maxlogit', 'randlogit', 'cost'] to get different ways to compute visualizations. Particularly in the visualize_vgg.py, you can change the variable "load_weights" from ['random', 'trained', 'part', 'reverse', 'only'] to decide different weight loading methods.

Finally, the pre-trained VGG-16 net is based on https://www.cs.toronto.edu/~frossard/post/vgg16/. You first need to download the trained weights vgg16_weights.npz into the deepnet directory.

Reference

@InProceedings{pmlr-v80-nie18a,
  title = 	 {A Theoretical Explanation for Perplexing Behaviors of Backpropagation-based Visualizations},
  author = 	 {Nie, Weili and Zhang, Yang and Patel, Ankit},
  booktitle = 	 {Proceedings of the 35th International Conference on Machine Learning},
  year = 	 {2018},
}