• check_coverage.sh -- run coverage over test suite (via run_tests.py) and report unit test coverage.
• cleanup.sh -- remove development detritus, pyc files, etc.
• dset_viz_attacked_stargalaxy.py -- visualize star-galaxy data attacked by an adversary.
• dset_viz_stargalaxy.py -- visualize star-galaxy data.
• fgsm_attack_stargalaxy.py -- script to run an FGSM attacker on the star-galaxy dataset.
• notebooks -- Jupyter notebooks for using the libraries here. See below for more information.
• ptlib/ -- lib dir for PyTorch code modules.
• requirements.txt -- packages needed to run everything.
• run_fgsm_attack_stargalaxy.sh -- bash script to wrap fgsm_attack_stargalaxy.py with convenience arguments.
• run_dset_viz_atacked_stargalaxy.sh -- bash script to wrap dset_viz_attacked_stargalaxy.py with convenience arguments.
• run_tests.py -- run all tests in tests/; generally best to avoid running this directly and instead run check_coverage.sh.
• run_update_todos.sh -- search code for "TODO" marks and drop them into the TODO.md file.
• run_vanilla_fashion.sh -- wrapper script for Fashion MNIST classifier.
• run_vanilla_stargalaxy.sh -- wrapper script for the star-galaxy classifier.
• tests/ -- lib dir for unittest modules.
• TODO.md -- programatically generated file with list of TODOs.
• vanilla_fashion.py -- script to run a Fashion MNIST classifier.
• vanilla_stargalaxy.py -- script to run a star-galaxy classifier.

Tensors are stored using PyTorch convnetions - [Number, Channel, Width, Height]

Images courtesy of Joao Caldeira (provenance unknown to me, but likely known to him). HDF5 data is available here:

https://github.com/gnperdue/RandomData/tree/master/StarGalaxy

e.g.

wget https://raw.githubusercontent.com/gnperdue/RandomData/master/StarGalaxy/stargalaxy_real_pt_test.hdf5
wget https://raw.githubusercontent.com/gnperdue/RandomData/master/StarGalaxy/stargalaxy_real_pt_train.hdf5
wget https://raw.githubusercontent.com/gnperdue/RandomData/master/StarGalaxy/stargalaxy_real_pt_valid.hdf5

NOTE -- the images stored in that Git repo are stored as uint8, but the code expects floating point numbers. There is a script in the RandomData/StarGalaxy package on GitHub for doing the conversion using the hdf5_manipulator package of utilities. (We are not storing the floats version in GitHub because those files are too large -- even the uint8 files are running up against their guidelines.)

HDF5 data is available here:

https://github.com/gnperdue/RandomData/tree/master/hdf5

e.g.

wget https://raw.githubusercontent.com/gnperdue/RandomData/master/hdf5/fashion_test.hdf5
wget https://raw.githubusercontent.com/gnperdue/RandomData/master/hdf5/fashion_train.hdf5

(Use 'test' for 'validation' -- no real 'test' file offered for this dset because it is fashion MNIST for crying out loud...)

The following notebooks provide a demonstration of the API. In order, they are:

1. star_galaxy.ipynb -- train a simple "star-galaxy" classifier and store the trained model locally.
2. attack_star_galaxy_fgsm.ipynb -- run the Fast Gradient Sign Method attack on the star-galaxy dataset for handful of attack strengths, using the model trained in the previous notebook, and save the attacked images in local HDF5 files.
3. visualiz_fgsm_attack_star_galaxy.ipynb -- visualize the "attacked" images for a given attack strength and save copies to a PDF for later viewing.