The expanded size of the tensor (100) must match the existing size (3) at non-singleton dimension 1. about pytorch-tvmisc HOT 5 CLOSED

Thank you for sharing your experience. Which notebook are you using and where did you need to add it?

from pytorch-tvmisc.

https://github.com/t-vi/pytorch-tvmisc/blob/master/wasserstein-distance/Pytorch_Wasserstein.ipynb

I'm testing it on histogram distance. but I'm not sure it could help unless I found a way to make it work on minibatches. right now the value is on all the stack as I see.

from pytorch-tvmisc.

The numerical stability of the batch approach there is not so good. I did implement a better way at some point but never got around to publishing it.

from pytorch-tvmisc.

maybe I can test it on a single minibatch tho, what I have seen until now that it is very slow funciton and maybe need to be implemented differently such as converting most of the functions into kernels.

from pytorch-tvmisc.

So the updated notebook explicitly handles batches. :)

from pytorch-tvmisc.