Package for calculating GAN metrics using Pytorch
Project description
Pytorch Implementation of Common GAN metrics
Notes
The FID implementation is inspired from pytorch-fid.
This repository is developed for personal research. If you think this package can also benefit your life, please feel free to open issues.
Install
pip install pytorch-gan-metrics
Feature
- Currently, this package supports following metrics:
- Inception Score (IS)
- Fréchet Inception Distance (FID)
- The computation processes of IS and FID are integrated to avoid multiple forward propagations.
- Support reading image on the fly to avoid out of memory especially for large scale images.
- Support computation on GPU to speed up some cpu operations such as
np.cov
andscipy.linalg.sqrtm
.
Reproducing Results of Official Implementations on CIFAR-10
Train IS | Test IS | Train(50k) vs Test(10k) FID |
|
---|---|---|---|
Official | 11.24±0.20 | 10.98±0.22 | 3.1508 |
pytorch-gan-metrics | 11.26±0.14 | 10.96±0.35 | 3.1518 |
pytorch-gan-metricsuse_torch=True |
11.26±0.15 | 10.96±0.19 | 3.1509 |
The results are slightly different from official implementations due to the framework difference between PyTorch and TensorFlow.
Documentation
Prepare Statistics for FID
- Download precalculated statistics or
- Calculate statistics for your custom dataset using command line tool
python -m pytorch_gan_metrics.calc_fid_stats --path path/to/images --output path/to/statistics.npz
See calc_fid_stats.py for implementation details.
Inception Features
When using pytorch_gan_metrics
to get IS or FID, the InceptionV3
will be loaded into torch.device('cuda:0')
if GPU is availabel; Otherwise, it uses cpu
to calculate inception features.
Using torch.Tensor
as images
- Prepare images in type
torch.float32
with shape[N, 3, H, W]
and normalized to[0,1]
.from pytorch_gan_metrics import (get_inception_score, get_fid, get_inception_score_and_fid) images = ... # [N, 3, H, W] assert 0 <= images.min() and images.max() <= 1 # Inception Score IS, IS_std = get_inception_score(images) # Frechet Inception Distance FID = get_fid(images, 'path/to/statistics.npz') # Inception Score & Frechet Inception Distance (IS, IS_std), FID = get_inception_score_and_fid( images, 'path/to/statistics.npz')
Using PyTorch DataLoader to Provide Images
- Use
pytorch_gan_metrics.ImageDataset
to collect images on disk or use customtorch.utils.data.Dataset
.from pytorch_gan_metrics import ImageDataset dataset = ImageDataset(path_to_dir, exts=['png', 'jpg']) loader = DataLoader(dataset, batch_size=50, num_workers=4)
- It is possible to wrap a generative model in a dataset to support generating images on the fly. Remember to set
num_workers=0
to avoid copying models across multiprocess.class GeneratorDataset(Dataset): def __init__(self, G, z_dim): self.G = G self.z_dim = z_dim def __len__(self): return 50000 def __getitem__(self, index): return self.G(torch.randn(1, self.z_dim).cuda())[0] dataset = GeneratorDataset(G, z=128) loader = DataLoader(dataset, batch_size=50, num_workers=0)
- Calculate metrics
from pytorch_gan_metrics import (get_inception_score, get_fid, get_inception_score_and_fid) # Inception Score IS, IS_std = get_inception_score(loader) # Frechet Inception Distance FID = get_fid(loader, 'path/to/statistics.npz') # Inception Score + Frechet Inception Distance (IS, IS_std), FID = get_inception_score_and_fid( loader, 'path/to/statistics.npz')
Specify Images by a Directory Path
- Calculate metrics for images in the directory.
from pytorch_gan_metrics import ( get_inception_score_from_directory, get_fid_from_directory, get_inception_score_and_fid_from_directory) IS, IS_std = get_inception_score_from_directory('path/to/images') FID = get_fid_from_directory('path/to/images', 'path/to/statistics.npz') (IS, IS_std), FID = get_inception_score_and_fid_from_directory( 'path/to/images', 'path/to/statistics.npz')
Accelerating Matrix Computation by PyTorch
- Set
use_torch=True
when calling functionsget_*
such asget_inception_score
,get_fid
, etc. - WARNING when
use_torch=True
is used, the FID might benan
due to the unstable implementation of matrix sqrt. - This option is recommended to be used when evaluating generative models on a server which is equipped with high efficiency GPUs while the cpu frequency is low.
License
This implementation is licensed under the Apache License 2.0.
This implementation is derived from pytorch-fid, licensed under the Apache License 2.0.
FID was introduced by Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler and Sepp Hochreiter in "GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium", see https://arxiv.org/abs/1706.08500
The original implementation of FID is by the Institute of Bioinformatics, JKU Linz, licensed under the Apache License 2.0. See https://github.com/bioinf-jku/TTUR.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for pytorch_gan_metrics-0.5.0.tar.gz
Algorithm | Hash digest | |
---|---|---|
SHA256 | 8f0b0d3d523bd0c505bb01c815c0e1d184a2b1039f67a541ef7e1e12eff9ff1a |
|
MD5 | 473821fdf4450d50600f733e7df1cb4e |
|
BLAKE2b-256 | 0c5d28d52887db4abc52a4c9aea9564de1e65a87bf11c712398adf076eaef76b |
Hashes for pytorch_gan_metrics-0.5.0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 4ed18e1f0702873dc10b7fc59cb3285cb3b5323990ab01f5fed9f74ba30bab97 |
|
MD5 | 67cb5dacc9da80a07e547fd654adb1be |
|
BLAKE2b-256 | 997715f3bce8ff187ab30934e16d9f4f50066e92e1293ec261d5d92d8c10582b |