A Port of Fréchet Inception Distance (FID score) to PyTorch
Clone or download
mseitzer Merge pull request #5 from ss18/ss18/pr0: Fix Typos
Fixed typos:

respectivly -> respectively
representive -> representative
Latest commit 7475159 Jan 8, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
LICENSE Initial commit Feb 10, 2018
README.md Update README.md Sep 24, 2018
fid_score.py respectivly -> respectively Jan 7, 2019
inception.py Add fid score Feb 10, 2018


Fréchet Inception Distance (FID score) in PyTorch

This is a port of the official implementation of Fréchet Inception Distance to PyTorch. See https://github.com/bioinf-jku/TTUR for the original implementation using Tensorflow.

FID is a measure of similarity between two datasets of images. It was shown to correlate well with human judgement of visual quality and is most often used to evaluate the quality of samples of Generative Adversarial Networks. FID is calculated by computing the Fréchet distance between two Gaussians fitted to feature representations of the Inception network.

Further insights and an independent evaluation of the FID score can be found in Are GANs Created Equal? A Large-Scale Study.

Note that the official implementation most likely gives slightly different scores, as it uses pretrained weights from Tensorflow 's Inception, not PyTorch's Inception as used here. If you report FID scores in your paper, and you want them to be comparable to FID scores reported in other papers, you should use the official Tensorflow implementation. You can still use this version if you want a quick FID estimate without installing Tensorflow.



  • python3
  • pytorch
  • torchvision
  • numpy
  • scipy

To compute the FID score between two datasets, where images of each dataset are contained in an individual folder:

./fid_score.py path/to/dataset1 path/to/dataset2

To run the evaluation on GPU, use the flag --gpu N, where N is the index of the GPU to use.

Using different layers for feature maps

In difference to the official implementation, you can choose to use a different feature layer of the Inception network instead of the default pool3 layer. As the lower layer features still have spatial extent, the features are first global average pooled to a vector before estimating mean and covariance.

This might be useful if the datasets you want to compare have less than the otherwise required 2048 images. Note that this changes the magnitude of the FID score and you can not compare them against scores calculated on another dimensionality. The resulting scores might also no longer correlate with visual quality.

You can select the dimensionality of features to use with the flag --dims N, where N is the dimensionality of features. The choices are:

  • 64: first max pooling features
  • 192: second max pooling featurs
  • 768: pre-aux classifier features
  • 2048: final average pooling features (this is the default)


This implementation is 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 is by the Institute of Bioinformatics, JKU Linz, licensed under the Apache License 2.0. See https://github.com/bioinf-jku/TTUR.