Skip to content

Latest commit

 

History

History
34 lines (26 loc) · 1.51 KB

pytorch_fid.md

File metadata and controls

34 lines (26 loc) · 1.51 KB
Raw

CleanFID vs PyTorch-FID (legacy)

PyTorch-legacy-FID refers to the popular PyTorch implementation of FID here code

Here we document the differences between CleanFID and PyTorch-legacy-FID. Additionally, we outline the steps that can be used to verify that using the mode="legacy_pytorch" flag is equivalent to the popular PyTorch implementation .

The key details are listed below:

  • The Repository used weights that are ported by the authors of the implementation from the TensorFlow model model to PyTorch
  • The resizing function this implementation of FID uses the bilinear filter implemented by the PyTorch ( speficially torch.nn.functional.interpolate).

Getting sample images

First download zipped file containing 5000 images generated by a StyleGAN2 model and 5000 randomly sampled from the FFHQ dataset.

wget <to_be_hosted>.zip

Compare the FID scores

  • First compute the FID score using the CleanFID with

    from cleanfid import fid

score = fid.compare_folders("test_fake/", "test_real/", mode="legacy_pytorch")

    The above gets a score of 8.966867513671616

  • Next compute the FID using the code

    pip install pytorch-fid

python -m pytorch_fid docs/test_fake/ docs/test_real/ \
        --device "cuda" --batch-size 128

    The above gets a score of 8.966869432185263