Add Perceptual Path Length

[jxtngx]

🚀 Feature

Add Perceptual Path Length (PPL) as described in this paper by NVIDIA:

A Style-Based Generator Architecture for Generative Adversarial Networks

Motivation

Add additional metrics to support the evaluation of generative models.

Pitch

Authors from Snap, ETH Zurich, and Carnegie Mellon used PyTorch Lightning in the following paper, but opted to use a metrics library other than torchmetrics. The only metric the authors needed that does not exist in torchmetrics is PPL.

Autodecoding Latent 3D Diffusion Models

[github-actions]

Hi! thanks for your contribution!, great first issue!

[rasbt]

Sounds great! One thing to add here is that we probably don't want to name it PPL (but maybe PerceptualPathLength), otherwise it will be confused with the popular perplexity metric

[SkafteNicki]

Hi @JustinGoheen, thanks for raising this issue.
We should definitly have this in torchmetrics. I try to look into it as soon as possible.
And agree with @rasbt that is should be PerceptualPathLength and not PPL. Usually we try to stick with long named, for example we use MeanSquaredError instead of MSE even if everyone in the universe agrees what the abbreviation stands for.