0.2.5

offer old resnet blocks from original ddpm paper, in case convnexts a…

…re unsuitable for generative work

0.2.4

extra norm on the queries as well for linear attention, plus extra ou…

…tput layernorm on the attention block, from sandwich norm and normformer papers

0.2.3

offer way to turn off sparse linear attention

0.2.2

give an initial conv

0.2.2

0.2.1

0.2.1

0.2.0

keep time embedding dimension as dim x 4, as openai does

0.1.6

automate more

0.1.5

offer way to load from last checkpoint, with `trainer.load(-1)`

0.1.2

no classifier free guidance if unet is not conditioned

0.1.1

small padding for the sampled grid of videos