Top-Down Networks

Code for reproducing the experiments of the paper Top-Down Networks: A coarse-to-fine reimagination of CNNs. In case of any bugs/improvements, please reach Ioannis Lelekas ([email protected]).

Requirements and Dependencies:

• Python >= 3.6
• NVIDIA GPU

We advise Anaconda for quick installation of all requirements and dependencies. Simply use the provided requirements.yml and conda env create -f requirements.yml.

Organization:

This repository is organized as follows:

./                        #root directory of repository

  data/                   #folder for downloading imagenette

  notebooks/              #notebooks for visualizing results

  src/                    #source code
    lib/                  #helper functions
    models/               #network architectures
    scripts/              #scripts for running the experiments

  output/                 #output folder; generated after running the code
    adversarial/          #extracted adversarial attacks
    gradcam/              #gradcam heatmaps
    graphs/               #training curves
    history/              #training history (loss, acc, learning curve)
    models/               #models checkpoints
    output/               #generated output from training
    trained_weights/      #trained weights for models

Usage:

All scripts for running the experiments are in src/scripts/. Command line inputs are given as comments within the scripts.

You may run the adversarial robustness (run_adversarial_attacks.py) and the localization experiment (run_gradcam.py) using models with pretrained weights. To this goal, download (link) and extract trained_weights.zip and then place it under src/output, as shown in the repository organization section.

Otherwise, you need to train models from scratch using the corresponding scripts (src.scripts.run_BUvsTD.py and src.scripts.run_imagenette.py).

The MD5 checksum of trained_weights.zip: 825009852499cb03781e45b905129615

Citation:

For citing Top-Down Networks, please use the following:

@inproceedings{lelekas2020,
  title={Top-Down Networks: A coarse-to-fine reimagination of CNNs},
  author={I Lelekas and N Tomen and SL Pintea and JC van Gemert},
  booktitle={CVPR 2020 Workshop on Deep Vision},
  year={2020}
}

Acknowledgements:

• https://github.com/BIGBALLON/cifar-10-cnn
• https://github.com/bethgelab/foolbox
• https://github.com/eclique/keras-gradcam
• https://github.com/koshian2/PCAColorAugmentation