/diverse_feature_vis

Code for Cadena et. al 2018 ECCV Diverse feature visualizations reveal invariances in early layers of deep neural networks.

Primary LanguagePythonMIT LicenseMIT

Diverse Feature Visualizations

This is a Python3 / Tensorflow implementation of the methods proposed in the paper:

Diverse feature visualizations reveal invariances in early layers of deep neural networks, by Santiago Cadena, Marissa Weis, Leon Gatys, Matthias Bethge, and Alexander Ecker.

Take a look at the two sample notebooks for the Diverse Visualizations of a paricular feature map of VGG19, and the shift-invariance test propsed in the paper.

Setup

To run this code you need the following:

  • Python3
  • Matplotlib
  • Tensorflow
  • Download the checkpoint weights of the normalized VGG network here (80MB), as well as the pixelcnn++ here or here if the later is broken (656MB), and store them in the networks/ folder

Our code uses the open-AI implementation of PixelCNN++ that can be found here.

Citation

If you find our code useful please cite us in your work:

@article{cadena2018diverse,
  title={Diverse feature visualizations reveal invariances in early layers of deep neural networks},
  author={Cadena, Santiago A and Weis, Marissa A and Gatys, Leon A and Bethge, Matthias and Ecker, Alexander S},
  journal={arXiv preprint arXiv:1807.10589},
  year={2018}
}