fix dimensions: the codebook must look at data by taking each time fr…

[wesbz]

…ame individually.
In SoundStream article: "This vector quantizer learns a
codebook of N vectors to encode each D-dimensional frame
of enc(x)."

[lucidrains]

@wesbz ohh, i'm actually operating with the input having dimensions batch x duration x features (see readme examples)

did you prefer to have it the other way around? (durations and features transposed)

[wesbz]

Oh! Alright, my bad, then!
Yes, it makes more sense to me to have batch × channels × duration ^^
Or I can just take that into account and correct it in my SoundStream implem ;) I'll just try that, do not close the PR yet, please ;)

[lucidrains]

@wesbz I can introduce a channel_last which can be set to False and do the transpose for you, if you wish

[wesbz]

This sounds perfect to me!

[lucidrains]

@wesbz ok done! https://github.com/lucidrains/vector-quantize-pytorch/releases/tag/0.3.8