StyleGAN2-ADA pytorch版本

资料仓库  收藏
0 / 618

StyleGAN2-ADA - Official PyTorch implementation

NVIDIA Research Projects

** Last update: Aug 3, 2021



StyleGAN2-ADA — Official PyTorch implementation

Teaser image

Training Generative Adversarial Networks with Limited Data
Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, Timo Aila

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.

Release notes

This repository is a faithful reimplementation of StyleGAN2-ADA in PyTorch, focusing on correctness, performance, and compatibility.


  • Full support for all primary training configurations.
  • Extensive verification of image quality, training curves, and quality metrics against the TensorFlow version.
  • Results are expected to match in all cases, excluding the effects of pseudo-random numbers and floating-point arithmetic.


  • Training is typically 5%–30% faster compared to the TensorFlow version on NVIDIA Tesla V100 GPUs.
  • Inference is up to 35% faster in high resolutions, but it may be slightly slower in low resolutions.
  • GPU memory usage is comparable to the TensorFlow version.
  • Faster startup time when training new networks (<50s), and also when using pre-trained networks (<4s).
  • New command line options for tweaking the training performance.


  • Compatible with old network pickles created using the TensorFlow version.

  • New ZIP/PNG based dataset format for maximal interoperability with existing 3rd party tools.

  • TFRecords datasets are no longer supported — they need to be converted to the new format.

  • New JSON-based format for logs, metrics, and training curves.

  • Training curves are also exported in the old TFEvents format if TensorBoard is installed.

  • Command line syntax is mostly unchanged, with a few exceptions (e.g.,

  • Comparison methods are not supported (--cmethod, --dcap, --cfg=cifarbaseline, --aug=adarv)

  • Truncation is now disabled by default.

Data repository

Path Description
stylegan2-ada-pytorch 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
    ├ ffhq.pkl FFHQ at 1024x1024, trained using original StyleGAN2
    ├ metfaces.pkl MetFaces at 1024x1024, transfer learning from FFHQ 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
    ├ brecahad.pkl BreCaHAD at 512x512, trained from scratch using ADA
    ├ paper-fig7c-training-set-sweeps Models used in Fig.7c (sweep over training set size)
    ├ paper-fig11a-small-datasets Models used in Fig.11a (small datasets & transfer learning)
    ├ paper-fig11b-cifar10 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


  • Linux and Windows are supported, but we recommend Linux for performance and compatibility reasons.
  • 1–8 high-end NVIDIA GPUs with at least 12 GB of memory. We have done all testing and development using NVIDIA DGX-1 with 8 Tesla V100 GPUs.
  • 64-bit Python 3.7, PyTorch 1.7.1, and CUDA toolkit 11.0 or newer. Use CUDA toolkit 11.1 or later with RTX 3090. See for PyTorch install instructions.
  • Python libraries: pip install click requests tqdm pyspng ninja imageio-ffmpeg==0.4.3. We use the Anaconda3 2020.11 distribution which installs most of these by default.
  • Docker users: use the provided Dockerfile to build an image with the required library dependencies.

The code relies heavily on custom PyTorch extensions that are compiled on the fly using NVCC. On Windows, the compilation requires Microsoft Visual Studio. We recommend installing Visual Studio Community Edition and adding it into PATH using "C:\Program Files (x86)\Microsoft Visual Studio\<version>\Community\VC\Auxiliary\Build\vcvars64.bat"</version>.

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 --outdir=out --trunc=1 --seeds=85,265,297,849 \

# Generate uncurated MetFaces images with truncation (Fig.12 upper left)
python --outdir=out --trunc=0.7 --seeds=600-605 \

# Generate class conditional CIFAR-10 images (Fig.17 left, Car)
python --outdir=out --seeds=0-35 --class=1 \

# Style mixing example
python --outdir=out --rows=85,100,75,458,1500 --cols=55,821,1789,293 \

Outputs from the above commands are placed under out/*.png, controlled by --outdir. Downloaded network pickles are cached under $HOME/.cache/dnnlib, which can be overridden by setting the DNNLIB_CACHE_DIR environment variable. The default PyTorch extension build directory is $HOME/.cache/torch_extensions, which can be overridden by setting TORCH_EXTENSIONS_DIR.

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

docker build --tag sg2ada:latest .
./ python3 --outdir=out --trunc=1 --seeds=85,265,297,849 \

Note: The Docker image requires NVIDIA driver release r455.23 or later.

Legacy networks: The above commands can load most of the network pickles created using the previous TensorFlow versions of StyleGAN2 and StyleGAN2-ADA. However, for future compatibility, we recommend converting such legacy pickles into the new format used by the PyTorch version:

python \
    --source= \

Projecting images to latent space

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

python --outdir=out --target=~/mytargetimg.png \

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

python --outdir=out --projected_w=out/projected_w.npz \

Using networks from Python

You can use pre-trained networks in your own Python code as follows:

with open('ffhq.pkl', 'rb') as f:
    G = pickle.load(f)['G_ema'].cuda()  # torch.nn.Module
z = torch.randn([1, G.z_dim]).cuda()    # latent codes
c = None                                # class labels (not used in this example)
img = G(z, c)                           # NCHW, float32, dynamic range [-1, +1]

The above code requires torch_utils and dnnlib to be accessible via PYTHONPATH. It does not need source code for the networks themselves — their class definitions are loaded from the pickle via torch_utils.persistence.

The pickle contains three networks. 'G' and 'D' are instantaneous snapshots taken during training, and 'G_ema' represents a moving average of the generator weights over several training steps. The networks are regular instances of torch.nn.Module, with all of their parameters and buffers placed on the CPU at import and gradient computation disabled by default.

The generator consists of two submodules, G.mapping and G.synthesis, that can be executed separately. They also support various additional options:

w = G.mapping(z, c, truncation_psi=0.5, truncation_cutoff=8)
img = G.synthesis(w, noise_mode='const', force_fp32=True)

Please refer to,, and for further examples.

Preparing datasets

Datasets are stored as uncompressed ZIP archives containing uncompressed PNG files and a metadata file dataset.json for labels.

Custom datasets can be created from a folder containing images; see python --help for more information. Alternatively, the folder can also be used directly as a dataset, without running it through first, but doing so may lead to suboptimal performance.

Legacy TFRecords datasets are not supported — see below for instructions on how to convert them.


Step 1: Download the Flickr-Faces-HQ dataset as TFRecords.

Step 2: Extract images from TFRecords using from the TensorFlow version of StyleGAN2-ADA:

# Using from TensorFlow version at
python ../stylegan2-ada/ unpack \
    --tfrecord_dir=~/ffhq-dataset/tfrecords/ffhq --output_dir=/tmp/ffhq-unpacked

Step 3: Create ZIP archive using from this repository:

# Original 1024x1024 resolution.
python --source=/tmp/ffhq-unpacked --dest=~/datasets/

# Scaled down 256x256 resolution.
python --source=/tmp/ffhq-unpacked --dest=~/datasets/ \
    --width=256 --height=256

MetFaces: Download the MetFaces dataset and create ZIP archive:

python --source=~/downloads/metfaces/images --dest=~/datasets/

AFHQ: Download the AFHQ dataset and create ZIP archive:

python --source=~/downloads/afhq/train/cat --dest=~/datasets/
python --source=~/downloads/afhq/train/dog --dest=~/datasets/
python --source=~/downloads/afhq/train/wild --dest=~/datasets/

CIFAR-10: Download the CIFAR-10 python version and convert to ZIP archive:

python --source=~/downloads/cifar-10-python.tar.gz --dest=~/datasets/

LSUN: Download the desired categories from the LSUN project page and convert to ZIP archive:

python --source=~/downloads/lsun/raw/cat_lmdb --dest=~/datasets/ \
    --transform=center-crop --width=256 --height=256 --max_images=200000

python --source=~/downloads/lsun/raw/car_lmdb --dest=~/datasets/ \
    --transform=center-crop-wide --width=512 --height=384 --max_images=200000


Step 1: Download the BreCaHAD dataset.

Step 2: Extract 512x512 resolution crops using from the TensorFlow version of StyleGAN2-ADA:

# Using from TensorFlow version at
python extract_brecahad_crops --cropsize=512 \
    --output_dir=/tmp/brecahad-crops --brecahad_dir=~/downloads/brecahad/images

Step 3: Create ZIP archive using from this repository:

python --source=/tmp/brecahad-crops --dest=~/datasets/

Training new networks

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

python --outdir=~/training-runs --data=~/ --gpus=1 --dry-run
python --outdir=~/training-runs --data=~/ --gpus=1

The first command is optional; it validates the arguments, prints out the training configuration, and exits. The second command kicks off the actual training.

In this example, the results are saved to a newly created directory ~/training-runs/<id>-mydataset-auto1</id>, controlled by --outdir. The training exports network pickles (network-snapshot-<int>.pkl</int>) and example images (fakes<int>.png</int>) at regular intervals (controlled by --snap). For each pickle, it also evaluates FID (controlled by --metrics) and logs the resulting scores in metric-fid50k_full.jsonl (as well as TFEvents if TensorBoard is installed).

The name of the output directory reflects the training configuration. For example, 00000-mydataset-auto1 indicates that the base configuration was auto1, meaning that the hyperparameters were selected automatically for training on one GPU. The base configuration is controlled by --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.

The training configuration can be further customized with additional command line options:

  • --aug=noaug disables ADA.
  • --cond=1 enables class-conditional training (requires a dataset with labels).
  • --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/<name>/network-snapshot-<int>.pkl</int></name> resumes a previous training run.
  • --gamma=10 overrides R1 gamma. We recommend trying a couple of different values for each new dataset.
  • --aug=ada --target=0.7 adjusts ADA target value (default: 0.6).
  • --augpipe=blit enables pixel blitting but disables all other augmentations.
  • --augpipe=bgcfnc enables all available augmentations (blit, geom, color, filter, noise, cutout).

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

Expected training time

The total training time depends heavily on resolution, number of GPUs, dataset, desired quality, and hyperparameters. The following table lists expected wallclock times to reach different points in the training, measured in thousands of real images shown to the discriminator ("kimg"):

Resolution GPUs 1000 kimg 25000 kimg sec/kimg GPU mem CPU mem
128x128 1 4h 05m 4d 06h 12.8–13.7 7.2 GB 3.9 GB
128x128 2 2h 06m 2d 04h 6.5–6.8 7.4 GB 7.9 GB
128x128 4 1h 20m 1d 09h 4.1–4.6 4.2 GB 16.3 GB
128x128 8 1h 13m 1d 06h 3.9–4.9 2.6 GB 31.9 GB
256x256 1 6h 36m 6d 21h 21.6–24.2 5.0 GB 4.5 GB
256x256 2 3h 27m 3d 14h 11.2–11.8 5.2 GB 9.0 GB
256x256 4 1h 45m 1d 20h 5.6–5.9 5.2 GB 17.8 GB
256x256 8 1h 24m 1d 11h 4.4–5.5 3.2 GB 34.7 GB
512x512 1 21h 03m 21d 22h 72.5–74.9 7.6 GB 5.0 GB
512x512 2 10h 59m 11d 10h 37.7–40.0 7.8 GB 9.8 GB
512x512 4 5h 29m 5d 17h 18.7–19.1 7.9 GB 17.7 GB
512x512 8 2h 48m 2d 22h 9.5–9.7 7.8 GB 38.2 GB
1024x1024 1 1d 20h 46d 03h 154.3–161.6 8.1 GB 5.3 GB
1024x1024 2 23h 09m 24d 02h 80.6–86.2 8.6 GB 11.9 GB
1024x1024 4 11h 36m 12d 02h 40.1–40.8 8.4 GB 21.9 GB
1024x1024 8 5h 54m 6d 03h 20.2–20.6 8.3 GB 44.7 GB

The above measurements were done using NVIDIA Tesla V100 GPUs with default settings (--cfg=auto --aug=ada --metrics=fid50k_full). "sec/kimg" shows the expected range of variation in raw training performance, as reported in log.txt. "GPU mem" and "CPU mem" show the highest observed memory consumption, excluding the peak at the beginning caused by torch.backends.cudnn.benchmark.

In typical cases, 25000 kimg or more is needed to reach convergence, but the results are already quite reasonable around 5000 kimg. 1000 kimg is often enough for transfer learning, which tends to converge significantly faster. The following figure shows example convergence curves for different datasets as a function of wallclock time, using the same settings as above:

Training curves

Note: --cfg=auto serves as a reasonable first guess for the hyperparameters but 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.

Quality metrics

By default, automatically computes FID for each network pickle exported during training. We recommend inspecting metric-fid50k_full.jsonl (or TensorBoard) at regular intervals to monitor the training progress. When desired, the automatic computation can be disabled with --metrics=none to speed up the training slightly (3%–9%).

Additional quality metrics can also be computed after the training:

# Previous training run: look up options automatically, save result to JSONL file.
python --metrics=pr50k3_full \

# Pre-trained network pickle: specify dataset explicitly, print result to stdout.
python --metrics=fid50k_full --data=~/datasets/ --mirror=1 \

The first example looks up the training configuration and performs the same operation as if --metrics=pr50k3_full had been specified during training. The second example downloads a pre-trained network pickle, in which case the values of --mirror and --data must be specified explicitly.

Note that many of the metrics have a significant one-off cost when calculating them for the first time for a new dataset (up to 30min). 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. Execution time and GPU memory usage is reported for one NVIDIA Tesla V100 GPU at 1024x1024 resolution:

Metric Time GPU mem Description
fid50k_full 13 min 1.8 GB Fréchet inception distance[1] against the full dataset
kid50k_full 13 min 1.8 GB Kernel inception distance[2] against the full dataset
pr50k3_full 13 min 4.1 GB Precision and recall[3] againt the full dataset
is50k 13 min 1.8 GB Inception score[4] for CIFAR-10

In addition, the following metrics from the StyleGAN and StyleGAN2 papers are also supported:

Metric Time GPU mem Description
fid50k 13 min 1.8 GB Fréchet inception distance against 50k real images
kid50k 13 min 1.8 GB Kernel inception distance against 50k real images
pr50k3 13 min 4.1 GB Precision and recall against 50k real images
ppl2_wend 36 min 2.4 GB Perceptual path length[5] in W, endpoints, full image
ppl_zfull 36 min 2.4 GB Perceptual path length in Z, full paths, cropped image
ppl_wfull 36 min 2.4 GB Perceptual path length in W, full paths, cropped image
ppl_zend 36 min 2.4 GB Perceptual path length in Z, endpoints, cropped image
ppl_wend 36 min 2.4 GB Perceptual path length in W, endpoints, cropped image


  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


Copyright © 2021, NVIDIA Corporation. All rights reserved.

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


  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}


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.


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.