1.5.9

add ability to get straight through gradients, as well as use reinmax

1.5.8

handle variable sequence lengths for cross entropy commitment loss

1.5.7

revert due to https://github.com/lucidrains/vector-quantize-pytorch/i…

1.5.6

revert due to https://github.com/lucidrains/vector-quantize-pytorch/i…

1.5.5

should be using l2norm of input for cosine sim

1.5.4

the quantized embeddings should be selected after the EMA step, so th…

1.5.3

in cosine sim, update with the l2norm of the smoothed embed_avg over …

1.5.2

cleanup cosine sim

1.5.1

fix an issue with multi-headed codebooks and reduction in cluster siz…

1.5.0

for cosine sim, keep track of the exponential moving averages of the …