Add support for computing sFID metric inside feature_extractor_inceptionv3.py

toshas / torch-fidelity

High-fidelity performance metrics for generative models in PyTorch

Other

1.01k stars 66 forks source link

Add support for computing sFID metric inside feature_extractor_inceptionv3.py #40

Closed giuseppecartella closed 1 year ago

giuseppecartella commented 1 year ago

As stated in the paper https://arxiv.org/pdf/2103.03841v1.pdf the computation of sFID metric requires the extraction of the first 7 channels from mixed6/conv feature map. Currently the repo does not provide it. What about adding this feature? It could be then exploited inside torchmetrics package.

toshas commented 1 year ago

That would probably require extending https://github.com/toshas/torch-fidelity/blob/master/torch_fidelity/feature_extractor_inceptionv3.py . It seems doable, so please feel free to suggest a change via the pull request. I will consider it for the next release, but no guarantees. Also, I don't see mass adoption of sFID in papers - can you point me to where else it is used and makes a difference against FID/KID?