vq-vae-2-pytorch icon indicating copy to clipboard operation
vq-vae-2-pytorch copied to clipboard

using vq-vae-2 for scoring probability of an image

Open zoharbarzelay opened this issue 5 years ago • 5 comments
trafficstars

Thanks a lot for this repo! quick question: once i train on your image set - can i somehow then use the model to estimate the probability that a new image is from the same distribution of the training set?

basically i want to use VAE to learn a prior model over images..

Thanks for anyone who can help with this!

Best, Z.

zoharbarzelay avatar Jul 09 '20 12:07 zoharbarzelay

I reconstruct the image and I compare reconstructed with original using SSIM comparison

venergiac avatar Jul 09 '20 13:07 venergiac

I reconstruct the image and I compare reconstructed with original using SSIM comparison

Thanks for your answer! could you plz provide some more details? for instance, what are the different colors in the graph?

zoharbarzelay avatar Jul 09 '20 13:07 zoharbarzelay

I think you can use prior (like PixelSNAIL) and reconstruction loss of the decoder to calculate likelihood.

rosinality avatar Jul 09 '20 14:07 rosinality

I think you can use prior (like PixelSNAIL) and reconstruction loss of the decoder to calculate likelihood.

thanks for your answer! how would you suggest that i use PixelSNAIL?

zoharbarzelay avatar Jul 20 '20 11:07 zoharbarzelay

@zozobozo I think you can use probability estimates of PixelSNAIL that you can calculate by forward the code of the sample.

rosinality avatar Jul 20 '20 13:07 rosinality