Glow

Code for reproducing results in "Glow: Generative Flow with Invertible 1x1 Convolutions"

To use pretrained CelebA-HQ model, make your own manipulation vectors and run our interactive demo, check demo folder.

Requirements

Tensorflow (tested with v1.8.0)
Horovod (tested with v0.13.8) and (Open)MPI

Run

pip install -r requirements.txt

To setup (Open)MPI, check instructions on Horovod github page.

Download datasets

For small scale experiments, use MNIST/CIFAR-10 (directly downloaded by train.py using keras)

For larger scale experiments, the datasets used are in the Google Cloud locations https://storage.googleapis.com/glow-demo/data/{dataset_name}-tfr.tar. The dataset_names are below, we mention the exact preprocessing / downsampling method for a correct comparison of likelihood.

Quantitative results

imagenet-oord - 20GB. Unconditional ImageNet 32x32 and 64x64, as described in PixelRNN/RealNVP papers (we downloaded this processed version).
lsun_realnvp - 140GB. LSUN 96x96. Random 64x64 crops taken at processing time, as described in RealNVP.

Qualitative results

celeba - 4GB. CelebA-HQ 256x256 dataset, as described in Progressive growing of GAN's. For 1024x1024 version (120GB), use celeba-full-tfr.tar while downloading.
imagenet - 20GB. ImageNet 32x32 and 64x64 with class labels. Centre cropped, area downsampled.
lsun - 700GB. LSUN 256x256. Centre cropped, area downsampled.

To download and extract celeb for example, run

wget https://storage.googleapis.com/glow-demo/data/celeba-tfr.tar
tar -xvf celeb-tfr.tar

Change hps.data_dir in train.py file to point to the above folder (or use the --data_dir flag when you run train.py)

For lsun, since download can be quite big, you can instead follow the instructions in data_loaders/generate_tfr/lsun.py to generate the tfr file directly from LSUN images. church_outdoor will be the smallest category.

Simple Train with 1 GPU

Run wtih small depth to test

CUDA_VISIBLE_DEVICES=0 python train.py --depth 1

Train with multiple GPUs using MPI and Horovod

Run default training script with 8 GPUs:

mpiexec -n 8 python train.py

Ablation experiments

mpiexec -n 8 python train.py --problem cifar10 --image_size 32 --n_level 3 --depth 32 --flow_permutation [0/1/2] --flow_coupling [0/1] --seed [0/1/2] --learntop --lr 0.001

Pretrained models, logs and samples

wget https://storage.googleapis.com/glow-demo/logs/abl-[reverse/shuffle/1x1]-[add/aff].tar

CIFAR-10 Quantitative result

mpiexec -n 8 python train.py --problem cifar10 --image_size 32 --n_level 3 --depth 32 --flow_permutation 2 --flow_coupling 1 --seed 0 --learntop --lr 0.001 --n_bits_x 8

ImageNet 32x32 Quantitative result

mpiexec -n 8 python train.py --problem imagenet-oord --image_size 32 --n_level 3 --depth 48 --flow_permutation 2 --flow_coupling 1 --seed 0 --learntop --lr 0.001 --n_bits_x 8

ImageNet 64x64 Quantitative result

mpiexec -n 8 python train.py --problem imagenet-oord --image_size 64 --n_level 4 --depth 48 --flow_permutation 2 --flow_coupling 1 --seed 0 --learntop --lr 0.001 --n_bits_x 8

LSUN 64x64 Quantitative result

mpiexec -n 8 python train.py --problem lsun_realnvp --category [bedroom/church_outdoor/tower] --image_size 64 --n_level 3 --depth 48 --flow_permutation 2 --flow_coupling 1 --seed 0 --learntop --lr 0.001 --n_bits_x 8