/stylegan2-ada

StyleGAN2 with adaptive discriminator augmentation (ADA) - Official TensorFlow implementation

Primary LanguageJupyter NotebookOtherNOASSERTION

PyTorch notice

I’ll likely be moving all features from this repo to the PyTorch version. That will take time, so until then this remains the best repo to use for full features.

Need Help?

  • If you’re new to StyleGAN2-ADA and looking to get started, please check out this video series from a course Lia Coleman and I taught in October 2020.
  • Interested in contributing? Please submit PRs or discuss changes in the Artificial Images Slack channel

Edits made to this repo

  • Fakes .jpg: save yourself a ton of space with fakes during training saved as .jpg instead of .png
  • Default GPU-based configs: added configs to maximize GPU usage for 11GB. 24GB, and 48GB cards (use 11GB for 16GB cards)
  • "Complex" configs: Apply the cifar10 config to other datasets (good for complex/diverse datasets) Use 11gb-gpu-complex or 24gb-gpu-complex
  • Conditional dataset creation from folders: Make conditional datasets from folders of images using create_from image_folders or create_from image_folders_raw
  • Vertical Mirroring: use --mirrory=True to flip training set top to bottom
  • Set Initial Augmentation Strength: use --initstrength={float value} to set the initialized strength of augmentations (really helpful when restarting training)
  • Set Initial Kimg count: use --nkimg={int value} to set the initial kimg count (helpful with restarts)
  • Set learnig rate from CLI: use --lrate={float value} to override config learning rate
  • Enable Two Time-Scale Update Rule: use --ttur=True to double the discriminator learnig rate compared to the generator.
  • Interpolations methods: Multiple interpolation methods included in the generate.py script. Linear and Spherical interpolations between seeds, and random noise or circular loops;
  • Neighbor vectors: Fine-tune seed selections by looking at vectors near it. Included in the generate.py script
  • Use np vectors in interpolations (in addition to seed values) Use saved .npy or .npz files in interpolation metohds. Thanks @ekkolabs!
  • Flesh Digressions: @aydao’s circular constant layer script edited to work with ADA see aydao_flesh_digressions.py
  • Raw dataset creations: Taken from the @skyflynil repo, reduces the size of datasets dramatically. Use create_from images_raw and create_from image_folders_raw in dataset creation, and use --use-raw=True in training (False by default!)
  • align faces script: From @pbaylies, this script will align images better for projection.
  • top-k training: Improve generator training by only propagating gradients from images the discriminator was most unsure of: Sinha & Zhao. Use --topk={float value} to define the decay rate.
  • @aydao's config: Extra large config for huge datasets (>100k img)
  • pkls blending scripts: From @justinpinkney, this scripts used for blending 2 pkls at various resolutions
  • JPG export: From @arthurfdlr, use --jpg_quality={int value} to define the JPG quality export, keep default value to export as PNG

StyleGAN2 with adaptive discriminator augmentation (ADA)
— Official TensorFlow implementation

Teaser image

Training Generative Adversarial Networks with Limited Data
Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, Timo Aila
https://arxiv.org/abs/2006.06676

Abstract: Training generative adversarial networks (GAN) using too little data typically leads to discriminator overfitting, causing training to diverge. We propose an adaptive discriminator augmentation mechanism that significantly stabilizes training in limited data regimes. The approach does not require changes to loss functions or network architectures, and is applicable both when training from scratch and when fine-tuning an existing GAN on another dataset. We demonstrate, on several datasets, that good results are now possible using only a few thousand training images, often matching StyleGAN2 results with an order of magnitude fewer images. We expect this to open up new application domains for GANs. We also find that the widely used CIFAR-10 is, in fact, a limited data benchmark, and improve the record FID from 5.59 to 2.42.

For business inquiries, please contact researchinquiries@nvidia.com
For press and other inquiries, please contact Hector Marinez at hmarinez@nvidia.com

What's new

This repository supersedes the original StyleGAN2 with the following new features:

  • ADA: Significantly better results for datasets with less than ~30k training images. State-of-the-art results for CIFAR-10.
  • Mixed-precision support: ~1.6x faster training, ~1.3x faster inference, ~1.5x lower GPU memory consumption.
  • Better hyperparameter defaults: Reasonable out-of-the-box results for different dataset resolutions and GPU counts.
  • Clean codebase: Extensive refactoring and simplification. The code should be generally easier to work with.
  • Command line tools: Easily reproduce training runs from the paper, generate projection videos for arbitrary images, etc.
  • Network import: Full support for network pickles produced by StyleGAN and StyleGAN2. Faster loading times.
  • Augmentation pipeline: Self-contained, reusable GPU implementation of extensive high-quality image augmentations.
  • Bugfixes

External data repository

Path Description
stylegan2-ada Main directory hosted on Amazon S3
  ├  ada-paper.pdf Paper PDF
  ├  images Curated example images produced using the pre-trained models
  ├  videos Curated example interpolation videos
  └  pretrained Pre-trained models
    ├  metfaces.pkl MetFaces at 1024x1024, transfer learning from FFHQ using ADA
    ├  brecahad.pkl BreCaHAD at 512x512, trained from scratch using ADA
    ├  afhqcat.pkl AFHQ Cat at 512x512, trained from scratch using ADA
    ├  afhqdog.pkl AFHQ Dog at 512x512, trained from scratch using ADA
    ├  afhqwild.pkl AFHQ Wild at 512x512, trained from scratch using ADA
    ├  cifar10.pkl Class-conditional CIFAR-10 at 32x32
    ├  ffhq.pkl FFHQ at 1024x1024, trained using original StyleGAN2
    ├  paper-fig7c-training-set-sweeps All models used in Fig.7c (baseline, ADA, bCR)
    ├  paper-fig8a-comparison-methods All models used in Fig.8a (comparison methods)
    ├  paper-fig8b-discriminator-capacity All models used in Fig.8b (discriminator capacity)
    ├  paper-fig11a-small-datasets All models used in Fig.11a (small datasets, transfer learning)
    ├  paper-fig11b-cifar10 All models used in Fig.11b (CIFAR-10)
    ├  transfer-learning-source-nets Models used as starting point for transfer learning
    └  metrics Feature detectors used by the quality metrics

Requirements

  • Linux and Windows are supported, but we recommend Linux for performance and compatibility reasons.
  • 64-bit Python 3.6 or 3.7. We recommend Anaconda3 with numpy 1.14.3 or newer.
  • We recommend TensorFlow 1.14, which we used for all experiments in the paper, but TensorFlow 1.15 is also supported on Linux. TensorFlow 2.x is not supported.
  • On Windows you need to use TensorFlow 1.14, as the standard 1.15 installation does not include necessary C++ headers.
  • 1–8 high-end NVIDIA GPUs with at least 12 GB of GPU memory, NVIDIA drivers, CUDA 10.0 toolkit and cuDNN 7.5.
  • Docker users: use the provided Dockerfile to build an image with the required library dependencies.

The generator and discriminator networks rely heavily on custom TensorFlow ops that are compiled on the fly using NVCC. On Windows, the compilation requires Microsoft Visual Studio to be in PATH. We recommend installing Visual Studio Community Edition and adding it into PATH using "C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build\vcvars64.bat".

Getting started

Pre-trained networks are stored as *.pkl files that can be referenced using local filenames or URLs:

# Generate curated MetFaces images without truncation (Fig.10 left)
python generate.py --outdir=out --trunc=1 --seeds=85,265,297,849 \
    --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada/pretrained/metfaces.pkl

# Generate uncurated MetFaces images with truncation (Fig.12 upper left)
python generate.py --outdir=out --trunc=0.7 --seeds=600-605 \
    --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada/pretrained/metfaces.pkl

# Generate class conditional CIFAR-10 images (Fig.17 left, Car)
python generate.py --outdir=out --trunc=1 --seeds=0-35 --class=1 \
    --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada/pretrained/cifar10.pkl

Outputs from the above commands are placed under out/*.png. You can change the location with --outdir. Temporary cache files, such as CUDA build results and downloaded network pickles, will be saved under $HOME/.cache/dnnlib. This can be overridden using the DNNLIB_CACHE_DIR environment variable.

Docker: You can run the above curated image example using Docker as follows:

docker build --tag stylegan2ada:latest .
docker run --gpus all -it --rm -v `pwd`:/scratch --user $(id -u):$(id -g) stylegan2ada:latest bash -c \
    "(cd /scratch && DNNLIB_CACHE_DIR=/scratch/.cache python3 generate.py --trunc=1 --seeds=85,265,297,849 \
    --outdir=out --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada/pretrained/metfaces.pkl)"

Projecting images to latent space

To find the matching latent vector for a given image file, run:

python projector.py --outdir=out --target=targetimg.png \
    --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada/pretrained/ffhq.pkl

For optimal results, the target image should be cropped and aligned similar to the original FFHQ dataset. The above command saves the projection target out/target.png, result out/proj.png, latent vector out/dlatents.npz, and progression video out/proj.mp4. You can render the resulting latent vector by specifying --dlatents for python generate.py:

python generate.py --outdir=out --dlatents=out/dlatents.npz \
    --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada/pretrained/ffhq.pkl

Preparing datasets

Datasets are stored as multi-resolution TFRecords, i.e., the same format used by StyleGAN and StyleGAN2. Each dataset consists of multiple *.tfrecords files stored under a common directory, e.g., ~/datasets/ffhq/ffhq-r*.tfrecords

MetFaces: Download the MetFaces dataset and convert to TFRecords:

python dataset_tool.py create_from_images ~/datasets/metfaces ~/downloads/metfaces/images
python dataset_tool.py display ~/datasets/metfaces

BreCaHAD: Download the BreCaHAD dataset. Generate 512x512 resolution crops and convert to TFRecords:

python dataset_tool.py extract_brecahad_crops --cropsize=512 \
    --output_dir=/tmp/brecahad-crops --brecahad_dir=~/downloads/brecahad/images

python dataset_tool.py create_from_images ~/datasets/brecahad /tmp/brecahad-crops
python dataset_tool.py display ~/datasets/brecahad

AFHQ: Download the AFHQ dataset and convert to TFRecords:

python dataset_tool.py create_from_images ~/datasets/afhqcat ~/downloads/afhq/train/cat
python dataset_tool.py create_from_images ~/datasets/afhqdog ~/downloads/afhq/train/dog
python dataset_tool.py create_from_images ~/datasets/afhqwild ~/downloads/afhq/train/wild
python dataset_tool.py display ~/datasets/afhqcat

CIFAR-10: Download the CIFAR-10 python version. Convert to two separate TFRecords for unconditional and class-conditional training:

python dataset_tool.py create_cifar10 --ignore_labels=1 \
    ~/datasets/cifar10u ~/downloads/cifar-10-batches-py

python dataset_tool.py create_cifar10 --ignore_labels=0 \
    ~/datasets/cifar10c ~/downloads/cifar-10-batches-py

python dataset_tool.py display ~/datasets/cifar10c

FFHQ: Download the Flickr-Faces-HQ dataset as TFRecords:

pushd ~
git clone https://github.com/NVlabs/ffhq-dataset.git
cd ffhq-dataset
python download_ffhq.py --tfrecords
popd
python dataset_tool.py display ~/ffhq-dataset/tfrecords/ffhq

LSUN: Download the desired LSUN categories in LMDB format from the LSUN project page and convert to TFRecords:

python dataset_tool.py create_lsun --resolution=256 --max_images=200000 \
    ~/datasets/lsuncat200k ~/downloads/lsun/cat_lmdb

python dataset_tool.py display ~/datasets/lsuncat200k

Custom: Custom datasets can be created by placing all images under a single directory. The images must be square-shaped and they must all have the same power-of-two dimensions. To convert the images to multi-resolution TFRecords, run:

python dataset_tool.py create_from_images ~/datasets/custom ~/custom-images
python dataset_tool.py display ~/datasets/custom

Training new networks

In its most basic form, training new networks boils down to:

python train.py --outdir=~/training-runs --gpus=1 --data=~/datasets/custom --dry-run
python train.py --outdir=~/training-runs --gpus=1 --data=~/datasets/custom

The first command is optional; it will validate the arguments, print out the resulting training configuration, and exit. The second command will kick off the actual training.

In this example, the results will be saved to a newly created directory ~/training-runs/<RUNNING_ID>-custom-auto1 (controlled by --outdir). The training will export network pickles (network-snapshot-<KIMG>.pkl) and example images (fakes<KIMG>.png) at regular intervals (controlled by --snap). For each pickle, it will also evaluate FID by default (controlled by --metrics) and log the resulting scores in metric-fid50k_full.txt.

The name of the output directory (e.g., 00000-custom-auto1) reflects the hyperparameter configuration that was used. In this case, custom indicates the training set (--data) and auto1 indicates the base configuration that was used to select the hyperparameters (--cfg):

Base config Description
auto (default) Automatically select reasonable defaults based on resolution and GPU count. Serves as a good starting point for new datasets, but does not necessarily lead to optimal results.
stylegan2 Reproduce results for StyleGAN2 config F at 1024x1024 using 1, 2, 4, or 8 GPUs.
paper256 Reproduce results for FFHQ and LSUN Cat at 256x256 using 1, 2, 4, or 8 GPUs.
paper512 Reproduce results for BreCaHAD and AFHQ at 512x512 using 1, 2, 4, or 8 GPUs.
paper1024 Reproduce results for MetFaces at 1024x1024 using 1, 2, 4, or 8 GPUs.
cifar Reproduce results for CIFAR-10 (tuned configuration) using 1 or 2 GPUs.
cifarbaseline Reproduce results for CIFAR-10 (baseline configuration) using 1 or 2 GPUs.

The training configuration can be further customized with additional arguments. Common examples:

  • --aug=noaug disables ADA (default: enabled).
  • --mirror=1 amplifies the dataset with x-flips. Often beneficial, even with ADA.
  • --resume=ffhq1024 --snap=10 performs transfer learning from FFHQ trained at 1024x1024.
  • --resume=~/training-runs/<RUN_NAME>/network-snapshot-<KIMG>.pkl resumes where a previous training run left off.
  • --gamma=10 overrides R1 gamma. We strongly recommend trying out at least a few different values for each new dataset.

Augmentation fine-tuning:

  • --aug=ada --target=0.7 adjusts ADA target value (default: 0.6).
  • --aug=adarv selects the alternative ADA heuristic (requires a separate validation set).
  • --augpipe=blit limits the augmentation pipeline to pixel blitting only.
  • --augpipe=bgcfnc enables all available augmentations (blit, geom, color, filter, noise, cutout).
  • --cmethod=bcr enables bCR with small integer translations.

Please refer to python train.py --help for the full list.

Expected training time

The total training time depends heavily on the resolution, number of GPUs, desired quality, dataset, and hyperparameters. In general, the training time can be expected to scale linearly with respect to the resolution and inversely proportional with respect to the number of GPUs. Small datasets tend to reach their lowest achievable FID faster than larger ones, but the convergence is somewhat less predictable. Transfer learning tends to converge significantly faster than training from scratch.

To give a rough idea of typical training times, the following figure shows several examples of FID as a function of wallclock time. Each curve corresponds to training a given dataset from scratch using --cfg=auto with a given number of NVIDIA Tesla V100 GPUs:

Training curves

Please note that --cfg=auto only serves as a reasonable first guess for the hyperparameters — it does not necessarily lead to optimal results for a given dataset. For example, --cfg=stylegan2 yields considerably better FID for FFHQ-140k at 1024x1024 than illustrated above. We recommend trying out at least a few different values of --gamma for each new dataset.

Preparing training set sweeps

In the paper, we perform several experiments using artificially limited/amplified versions of the training data, such as ffhq30k, ffhq140k, and lsuncat30k. These are constructed by first unpacking the original dataset into a temporary directory with python dataset_tool.py unpack and then repackaging the appropriate versions into TFRecords with python dataset_tool.py pack. In the following examples, the temporary directories are created under /tmp and can be safely deleted afterwards.

# Unpack FFHQ images at 256x256 resolution.
python dataset_tool.py unpack --resolution=256 \
    --tfrecord_dir=~/ffhq-dataset/tfrecords/ffhq --output_dir=/tmp/ffhq-unpacked

# Create subset with 30k images.
python dataset_tool.py pack --num_train=30000 --num_validation=10000 --seed=123 \
    --tfrecord_dir=~/datasets/ffhq30k --unpacked_dir=/tmp/ffhq-unpacked

# Create amplified version with 140k images.
python dataset_tool.py pack --num_train=70000 --num_validation=0 --mirror=1 --seed=123 \
    --tfrecord_dir=~/datasets/ffhq140k --unpacked_dir=/tmp/ffhq-unpacked

# Unpack LSUN Cat images at 256x256 resolution.
python dataset_tool.py unpack --resolution=256 \
    --tfrecord_dir=~/datasets/lsuncat200k --output_dir=/tmp/lsuncat200k-unpacked

# Create subset with 30k images.
python dataset_tool.py pack --num_train=30000 --num_validation=10000 --seed=123 \
    --tfrecord_dir=~/datasets/lsuncat30k --unpacked_dir=/tmp/lsuncat200k-unpacked

Please note that when training with artifically limited/amplified datasets, the quality metrics (e.g., fid50k_full) should still be evaluated against the corresponding original datasets. This can be done by specifying a separate metric dataset for train.py and calc_metrics.py using the --metricdata argument. For example:

python train.py [OTHER_OPTIONS] --data=~/datasets/ffhq30k --metricdata=~/ffhq-dataset/tfrecords/ffhq

Reproducing training runs from the paper

The pre-trained network pickles (stylegan2-ada/pretrained/paper-fig*) reflect the training configuration the same way as the output directory names, making it straightforward to reproduce a given training run from the paper. For example:

# 1. AFHQ Dog
# paper-fig11a-small-datasets/afhqdog-mirror-paper512-ada.pkl
python train.py --outdir=~/training-runs --gpus=8 --data=~/datasets/afhqdog \
    --mirror=1 --cfg=paper512 --aug=ada

# 2. Class-conditional CIFAR-10
# pretrained/paper-fig11b-cifar10/cifar10c-cifar-ada-best-fid.pkl
python train.py --outdir=~/training-runs --gpus=2 --data=~/datasets/cifar10c \
    --cfg=cifar --aug=ada

# 3. MetFaces with transfer learning from FFHQ
# paper-fig11a-small-datasets/metfaces-mirror-paper1024-ada-resumeffhq1024.pkl
python train.py --outdir=~/training-runs --gpus=8 --data=~/datasets/metfaces \
    --mirror=1 --cfg=paper1024 --aug=ada --resume=ffhq1024 --snap=10

# 4. 10k subset of FFHQ with ADA and bCR
# paper-fig7c-training-set-sweeps/ffhq10k-paper256-ada-bcr.pkl
python train.py --outdir=~/training-runs --gpus=8 --data=~/datasets/ffhq10k \
    --cfg=paper256 --aug=ada --cmethod=bcr --metricdata=~/ffhq-dataset/tfrecords/ffhq

# 5. StyleGAN2 config F
# transfer-learning-source-nets/ffhq-res1024-mirror-stylegan2-noaug.pkl
python train.py --outdir=~/training-runs --gpus=8 --data=~/ffhq-dataset/tfrecords/ffhq \
    --res=1024 --mirror=1 --cfg=stylegan2 --aug=noaug --metrics=fid50k

Notes:

  • You can use fewer GPUs than shown in the above examples. This will only increase the training time — it will not affect the quality of the results.
  • Example 3 specifies --snap=10 to export network pickles more frequently than usual. This is recommended, because transfer learning tends to yield very fast convergence.
  • Example 4 specifies --metricdata to evaluate quality metrics against the original FFHQ dataset, not the artificially limited 10k subset used for training.
  • Example 5 specifies --metrics=fid50k to evaluate FID the same way as in the StyleGAN2 paper (see below).

Quality metrics

By default, train.py will automatically compute FID for each network pickle. We strongly recommend inspecting metric-fid50k_full.txt at regular intervals to monitor the training progress. When desired, the automatic computation can be disabled with --metrics none to speed up the training.

Additional quality metrics can also be computed after the training:

# Previous training run: look up options automatically, save result to text file.
python calc_metrics.py --metrics=pr50k3_full \
    --network=~/training-runs/00000-ffhq10k-res64-auto1/network-snapshot-000000.pkl

# Pretrained network pickle: specify dataset explicitly, print result to stdout.
python calc_metrics.py --metrics=fid50k_full --metricdata=~/datasets/ffhq --mirror=1 \
    --network=https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada/pretrained/ffhq.pkl

The first example will automatically find training_options.json stored alongside the network pickle and perform the same operation as if --metrics pr50k3_full had been specified during training. The second example will download a pre-trained network pickle, in which case the values of --mirror and --metricdata have to be specified explicitly.

Note that many of the metrics have a significant one-off cost (up to an hour or more) when they are calculated for the first time using a given dataset. Also note that the evaluation is done using a different random seed each time, so the results will vary if the same metric is computed multiple times.

We employ the following metrics in the ADA paper. The expected execution times correspond to using one Tesla V100 GPU at 1024x1024 and 256x256 resolution:

Metric 1024x1024 256x256 Description
fid50k_full 15 min 5 min Fréchet inception distance[1] against the full dataset.
kid50k_full 15 min 5 min Kernel inception distance[2] against the full dataset.
pr50k3_full 20 min 10 min Precision and recall[3] againt the full dataset.
is50k 25 min 5 min Inception score[4] for CIFAR-10.

In addition, all metrics that were used in the StyleGAN and StyleGAN2 papers are also supported for backwards compatibility:

Legacy: StyleGAN2 1024x1024 Description
fid50k 15 min Fréchet inception distance against 50k real images.
kid50k 15 min Kernel inception distance against 50k real images.
pr50k3 20 min Precision and recall against 50k real images.
ppl2_wend 40 min Perceptual path length[5] in W at path endpoints against full image.
Legacy: StyleGAN 1024x1024 Description
ppl_zfull 40 min Perceptual path length in Z for full paths against cropped image.
ppl_wfull 40 min Perceptual path length in W for full paths against cropped image.
ppl_zend 40 min Perceptual path length in Z at path endpoints against cropped image.
ppl_wend 40 min Perceptual path length in W at path endpoints against cropped image.
ls 10 hrs Linear separability[5] with respect to CelebA attributes.

References:

  1. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium, Heusel et al. 2017
  2. Demystifying MMD GANs, Bińkowski et al. 2018
  3. Improved Precision and Recall Metric for Assessing Generative Models, Kynkäänniemi et al. 2019
  4. Improved Techniques for Training GANs, Salimans et al. 2016
  5. A Style-Based Generator Architecture for Generative Adversarial Networks, Karras et al. 2018

License

Copyright © 2020, NVIDIA Corporation. All rights reserved.

This work is made available under the Nvidia Source Code License.

Citation

@inproceedings{Karras2020ada,
  title     = {Training Generative Adversarial Networks with Limited Data},
  author    = {Tero Karras and Miika Aittala and Janne Hellsten and Samuli Laine and Jaakko Lehtinen and Timo Aila},
  booktitle = {Proc. NeurIPS},
  year      = {2020}
}

Development

This is a research reference implementation and is treated as a one-time code drop. As such, we do not accept outside code contributions in the form of pull requests.

Acknowledgements

We thank David Luebke for helpful comments; Tero Kuosmanen and Sabu Nadarajan for their support with compute infrastructure; and Edgar Schönfeld for guidance on setting up unconditional BigGAN.