Skip to content
Reproducible code for Augmentation paper
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.



  • Python 3.6+
  • Pillow, matplotlib, numpy, pytorch==0.4.0, seaborn, torchvision


  • runs a full set of experiments on MNIST and save the results to the directory saved. Note: the default run take a long time (43 hours) to finish, since we're running for all 10 seeds.


    Currently, it executes the following experiments:

    1. Measure the difference between exact augmented objective and approximate objectives (on original images, 1st order approximation, 2nd order approximation).
    2. Measure the agreement and KL divergence between the predictions made by model trained on exact augmented objective and models trained on approximate objectives.
    3. Compute kernel target alignment for features from different transformations.
  • plots all the figures in the paper using the saved results from The figures are saved in the directory figs.

You can’t perform that action at this time.