GAN Metrics

Inception Score

Run python inception_score.py --dataroot "path/to/image/folder" to calculate IS

Because I use Pytorch ImageFolder class from torchvision.datasets, path/to/image/folder should contain folder of images. Please refer to the documentation for more detail.

path
└── to
    └── image
         └── folder
                └── class0
                        ├── 0.jpg
                        └── 1.jpg
                └── class1
                        ├── 0.jpg
                        └── 1.jpg

IS code is derived from sbarratt

Fréchet Inception Distance

Run python fid_score.py path/to/image/folder1 path/to/image/folder2 --gpu 0 to calculate FID. The folder1 and folder2 should contain images.

path
└── to
    └── image
         └── folder
                ├── 0.jpg
                └── 1.jpg

Option to calculate FID on multiple generated image folders. python fid_score.py path/to/image/folder path/to/generated/folder --gpu 0 --multiple_gen_dir. This command will output a summary file called fid.csv. Run this command to find the best FID epoch.

path
└── to
    └── generated
         └── folder
                └── epoch0
                        ├── 0.jpg
                        └── 1.jpg
                └── epoch1
                        ├── 0.jpg
                        └── 1.jpg

FID code is derived from mseitzer