0.7.8

fix potential bug with discriminator training in vqgan-vae

0.7.7a

do attention in float32 always

0.7.7

do attention in float32 always

0.7.6

lock codebook dimensions to 256, unless if overridden for something s…

…maller, as in improved vqgan paper

0.7.5

only apply gradient penalty every 4th step, like stylegan

0.7.4

fix issue with EMA class

0.7.3

remove all biases, given results from PaLM as well as hearing about @…

…borisdayma experiments

0.7.2

move all exponential moving average logic to separate class, add an E…

…MA for the vqgan-vae during training

0.7.1

cleanup

0.7.0

move to using causal convolutions through asymmetric padding