Vector Quantized VAEs

Vector Quantized VAEs are the first variational auto-encoders to be competitive with GANs in the quality of the generated images.

[@K3pukk4]

what a legend!

[@jonathanyang2359]

Thanks! I don't attend this institution, but this was an extremely clear lecture :)

[@LyoshaZebra]

Thanks for explaining that! Great job. Subscribed!

[@sdfrtyhfds]

do you train the pixel cnn on the same data and just not update the Vae weights while training?

[@davidmcallester4973]

yes, the vector quantization is held constant as the pixel CNN is trained.

[@sdfrtyhfds]

also, what if you skip the quantization during inference? would you still get images that make sense?

[@davidmcallester4973]

Do you mean "during generation"? During generation you can't skip the quantization because the pixel-CNN is defined to generate the quantized vectors (the symbols).

[@sdfrtyhfds]

@@davidmcallester4973 I guess that during generation it wouldn't make much sense, i was thinking more in the direction of interpolating smoothly between two different symbols.

[@bernhardbermeitinger8617]

Thank you for this video, however, please don't call your variable ŝ 😆 (or at least don't say it out loud)