Sliced Iterative Normalizing Flows

This repository is the official implementation of Sliced Iterative Normalizing Flows.

Requirements

To install requirements:

pip install -r requirements.txt

Training

To train SIG and GIS, run the following commands:

python SIG.py --dataset DATASET --seed SEED --save SAVING_ADDRESS   
python GIS.py --dataset DATASET --seed SEED --save SAVING_ADDRESS

Evaluation

The FID score of SIG samples can also be evaluated on the fly with the "--evaluateFID" argument.

Results

SIG achieves the following performance on :

FID score

| MNIST | Fashion | CIFAR10 | CelebA | | ----------------- | ----------------- | ----------------- | ----------------- | | 4.5 | 13.7 | 66.5 | 37.3 |

OoD detection (AUROC on models trained on FashionMNIST)

| MNIST | OMNIGLOT | FMNIST-hflip | FMNIST-vflip | | ----------------- | ----------------- | ----------------- | ----------------- | | 0.990 | 0.993 | 0.631 | 0.821 |