As stated previously, this implementation does not generate similar scores as the original tensorflow implementation. I recommend everyone to use mseitzer's implementation: https://github.com/mseitzer/pytorch-fid
This implementation contains an implementation of the FID score in pytorch using a pre-trained InceptionV3 network.
NOTE This is not the official implementation of FID. View Two time-scale update rule for training GANs for the official implementation of FID. This implementation uses Pytorch pre-trained InceptionV3 network which is the same as the tensorflow pre-trained network.
FID is a performance metric to evaluate the similarity between two dataset of images. It was introduced by the paper "Two time-scale update rule for training GANs". It is shown to correlate well to human evaluation of image quality, and it is able to detect intra-class mode collapse.
This implementation uses a different InceptionV3 model from what used in the original implementation. Not confirmed, but we believe pytorch uses the InceptionV3 network from 2016-08-28, while the original FID implementation uses a network from 2015-12-05.
Original implementation InceptionV3 model:
- 2015-12-05, code reference
- python3 (not tested for python2)
- torch & torchvision 1.0 (might work for 0.4.0+)
- numpy
- scipy
- opencv
To compute FID between to datasets, use the following command:
python fid.py --path1 path/to/real/data --path2 path/to/fake/data --batch-size 8The script will use all .png and .jpg files in the directories. It assumes the following:
The images are
- unsigned int 8 between 0-255, or float32 between 0-1.
- all images are of same image shape
To compute InceptionV3 activations we:
- reshape images to (299,299). This is done in the official FID implementation
- normalize image similarly to [-1, 1]
- use network.eval() to use the running average for batch normalization.
- Port the InceptionV3 network used from the original implementation instead of using pytorch default.