0

I am using stable diffusion code to train a diffusion model. Their paper mentioned FID scores assessed on 5000 samples using 100 DDIM steps. The git repository does not have the implementation for how they computed the FID score over 5000 samples. I plan on using the torch metrics FID to compute it. What is not clear to me how to do it for random set of 5000 images in pytorch lightning run? My question is how to report FID score metric for my training and validation epochs?

I have never used generative AI metrics before.

bananagator
  • 551
  • 3
  • 8
  • 26
  • Could you modify that code to log that metric over the training iterations for Tensorboard? – paisanco Jul 02 '23 at 15:43
  • @paisanco is the FID computed between images and reconstructions or input latent space and reconstructed latent space? Is the FID score reported as a mean of metric generated for each batch or do you call the update function each batch and compute metric at the end of the epoch? The update and compute functions refer to torch metric way of computing FID score. See the example in the torchmetrics link i attached in the question. – bananagator Jul 02 '23 at 16:14

0 Answers0