CFID

We consider distance functions between conditional distributions functions. We focus on the Wasserstein metric and its Gaussian case known as the Frechet Inception Distance (FID).We develop conditional versions of these metrics, and analyze their relations. Then, we numerically compare the metrics inthe context of performance evaluation of conditional generative models. Our results show that the metrics are similar in classical models which are less susceptible to conditional collapse. But the conditional distances are more informative in modern unsupervised, semisupervised and unpaired models where learning the relations between the inputs and outputs is the main challenge.

CFID identifies the good models (paired), and yields bad scores to models that ignore their inputs (unpaired).

Project Page

Paper

Code

TensorFlow2 implementation of Conditional Frechet Inception Distance metric. For example, given a low-resolution (LR) image, ground-truth high-resolution image (HR) and some super-resolution model (SR). The CFID metric is able to measure the distance between HR and SR given LR. In comparison to classic Frechet Inception Distance (FID), CFID considers the input LR image. It measures the similarity between HR and SR with respect to the input image. Unlike FID, CFID requires paired (LR,HR) data for comparison.

We also introduce RFID metric as a condidate for measuring the distance between distributions. The corresponding graphical models describe the difference between the metrics:

FID, RFID, CFID

Note: The current software works well with TensorFlow 2.4.0

Conditional Frechet Inception Distance.
Michael Soloveitchik, Tzvi Diskin, Efrat Morin, Ami Wiesel.
The Hebrew University of Jerusalem, 2021.

Formulas

The CFID formula is similar to FID, simple and easy to implement. Given:

CFID defined as follows:

Example results

‘good’ models defined to be those which output corellate visually with the input. For example when the SR image could be donwsampled back to it’s LR input. CFID distinguish between ‘good’ and ‘bad’ models while the classic FID metric doesn’t. Most of the models that trained on paired data considered to be ‘good’. We provide comparasion of CFID with FID on ‘good’ and ‘bad’ models.

The ‘good’ models are: Pix2Pix and BiCycle-GAN (The were trained on paired data).

The ‘bad’ models are: Cycle-GAN and MUNIT (They were trained on un-paired data)

The models were trained on Celeb-A dataset

• Pix2Pix vs Cycle-GAN
• 
• biCycle-GAN vs MUNIT
• 

Prerequisites

• Windows, Linux or macOS
• Python 3
• CPU or NVIDIA GPU + CUDA CuDNN

Getting Started

Installation

• Clone this repo:

  git clone -b master --single-branch https://github.com/Michael-Soloveitchik/CFID.git
  cd CFID
  

• Install Tensorflow and dependencies from https://www.tensorflow.org/

Citation

If you find this useful for your research, please use the following.

@misc{soloveitchik2021conditional,
      title={Conditional Frechet Inception Distance}, 
      author={Michael Soloveitchik and Tzvi Diskin and Efrat Morin and Ami Wiesel},
      year={2021},
      eprint={2103.11521},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Acknowledgements

The authors are with The Hebrew University of Jerusalem, Israel . This research was funded by Center for Interdisciplinary Data Science (CIDR) in The Hebrew University of Jerusalem, Israel.

This site is open source. Improve this page.