Densely connected normalizing flows

This repository is the official implementation of NeurIPS 2021 paper Densely connected normalizing flows. Poster available here.

Setup

CUDA 11.1
Python 3.8

pip install -r requirements.txt
pip install -e .

Training

cd ./experiments/image

CIFAR-10:

python train.py --epochs 400 --batch_size 64 --optimizer adamax --lr 1e-3  --gamma 0.9975 --warmup 5000  --eval_every 1 --check_every 10 --dataset cifar10 --augmentation eta --block_conf 6 4 1 --layers_conf  5 6 20  --layer_mid_chnls 48 48 48 --growth_rate 10  --name DF_74_10

python train_more.py --model ./log/cifar10_8bit/densenet-flow/expdecay/DF_74_10 --new_lr 2e-5 --new_epochs 420

ImageNet32:

python train.py --epochs 20 --batch_size 64 --optimizer adamax --lr 1e-3  --gamma 0.95 --warmup 5000  --eval_every 1 --check_every 10 --dataset imagenet32 --augmentation eta --block_conf 6 4 1 --layers_conf  5 6 20  --layer_mid_chnls 48 48 48 --growth_rate 10  --name DF_74_10

python train_more.py --model ./log/imagenet32_8bit/densenet-flow/expdecay/DF_74_10 --new_lr 2e-5 --new_epochs 22

ImageNet64:

python train.py --epochs 10 --batch_size 32 --optimizer adamax --lr 1e-3  --gamma 0.95 --warmup 5000  --eval_every 1 --check_every 10 --dataset imagenet64 --augmentation eta --block_conf 6 4 1 --layers_conf  5 6 20  --layer_mid_chnls 48 48 48 --growth_rate 10  --name DF_74_10

python train_more.py --model ./log/imagenet64_8bit/densenet-flow/expdecay/DF_74_10 --new_lr 2e-5 --new_epochs 11

CelebA:

python train.py --epochs 50 --batch_size 32 --optimizer adamax --lr 1e-3  --gamma 0.95 --warmup 5000  --eval_every 1 --check_every 10 --dataset celeba --augmentation horizontal_flip --block_conf 6 4 1 --layers_conf  5 6 20  --layer_mid_chnls 48 48 48 --growth_rate 10  --name DF_74_10

python train_more.py --model ./log/celeba_8bit/densenet-flow/expdecay/DF_74_10 --new_lr 2e-5 --new_epochs 55

Note: Download instructions for ImageNet and CelebA are defined in denseflow/data/datasets/image/{dataset}.py

Evaluation

CIFAR-10:

python eval_loglik.py --model PATH_TO_MODEL --k 1000 --kbs 50

ImageNet32:

python eval_loglik.py --model PATH_TO_MODEL --k 200 --kbs 50

ImageNet64 and CelebA:

python eval_loglik.py --model PATH_TO_MODEL --k 200 --kbs 25

Model weights

Model weights are stored here.

Update Dec 2023. Note that our ImageNet models are trained on the publicly available version of the dataset at https://image-net.org

Samples generation

Generated samples are stored in PATH_TO_MODEL/samples

python eval_sample.py --model PATH_TO_MODEL

Note: PATH_TO_MODEL has to contain check directory.

ImageNet 32x32

ImageNet 64x64

CelebA

Acknowledgements

Significant part of this code benefited from SurVAE [1] code implementation, available under MIT license.

References

[1] Didrik Nielsen, Priyank Jaini, Emiel Hoogeboom, Ole Winther, and Max Welling. Survae flows: Surjections to bridge the gap between vaes and flows. InAdvances in Neural Information Processing Systems 33. Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020.

matejgrcic/DenseFlow