1.6.1

remove what reportedly does not work well

1.6.0

buy the conclusions of the new MIT paper with their solution for the …

1.5.19

fix cosine sim and multi-headed codebooks

1.5.18

if not training, no need for straight through calculation nor stochas…

1.5.17

temperature should also be applied to logits for sampling, unless if …

1.5.16

allow for temperature to be annealed

1.5.15

fix a bug with straight-through, thanks to @goldbird5

1.5.14

use numerically stable log in reinmax

1.5.12

default temperature to 1.

1.5.11

only quantize dropout if number of quantizers greater than 1