pytorch-fcd 1.0.1.dev1

Package for calculating Frechet CLIP Distance (FCD) using PyTorch

FCD_Score

Fréchet CLIP Distance Score: Fréchet CLIP Distance Score calculate Fréchet Distance based on features from CLIP image encoder.
FCD measure similarity between generated images and the dataset.

Installation

Install from pip:
pip install pytorch-fcd
Requirements:
ftfy
regex
numpy
pillow
scipy
torch>=1.7.1
torchvision>=0.8.2
clip

Usage

calculate and save statistics for datasest:
pytorch-fcd --model-name "ViT-B/32" --savepath path_to_image_dataset path_to_save_file

calculate FCD score for image dataset and generated image dataset:
pytorch-fcd --model-name "ViT-B/32" --paths path_to_image_dataset path_to_generated_image_dataset

Citing

If you use this repository in your research, consider citing it using the following Bibtex entry:
@misc{ShanYang2023FCD,
author={Shan Yang},
title={{pytorch-fcd: FCD Score for PyTorch}},
month={Nov},
year={2023},
note={Version 1.0.0.dev1},
howpublished={\url{https://github.com/syang-lab/FCD_Score}} }

License

This implementation is licensed under the Apache License 2.0.