"Training Normalizing Flows with the Information Bottleneck for Competitive Generative Classification" (2020)
https://arxiv.org/abs/2001.06448
- All configuration files for all experiments in the paper are contained in the directory 'experiments_configs'.
Training (for one example configuration):
python main.py train experiments_configs/cifar10/beta_ramp/beta_1.0000.ini
Testing:
python main.py test experiments_configs/cifar10/beta_ramp/beta_1.0000.ini`
- The cifar/mnist datasets should be downloaded automatically the first time it is run. For the OoD evaluation, tiny imagenet and quickdraw have to be downloaded separately.
To implement the INNs, we use of the FrEIA library (github.com/VLL-HD/FrEIA)
pip install git+https://github.com/VLL-HD/FrEIA.gitAdditional requirements:
- pytorch=1.4.0
- numpy=1.18.1
- matplotlib=3.1.3
- torchvision=0.2.2
(other versions will likely work too, but have not been tested)