mblondel/soft-dtw

Pytorch feature

Rachine opened this issue · 9 comments

Rachine commented

Hi !

Did you start working on the pytorch feature? Or should I start looking into it?
Thanks for sharing the code!

mblondel commented

@arthurmensch is also interested in this. I guess you two could collaborate!

sculyi commented

hi, i have the same problem, did you have any ideas?

amirkargar commented

Hi, did you guys have any progress on this?

mblondel commented

Closing since a PyTorch implementation is now available in https://github.com/arthurmensch/didyprog thanks to @arthurmensch.

  • 👍1
shaform commented

It seems like a Pytorch implementation is not really available in that repo arthurmensch/didyprog#5. So maybe this should be reopened?

mblondel commented

@shaform I agree. Does anyone want to tackle it? I won't be able to do it myself, since I don't have the use case myself. Here are the things to do:

  • Create a pytorch function class (https://pytorch.org/docs/stable/notes/extending.html). This should be pretty straightforward (cf. chainer class for inspiration).
  • Write a unit test that makes sure the gradient is correct.
  • Port chainer example in examples/ to pytorch.
lyprince commented

@mblondel Here is my implementation (https://github.com/lyprince/sdtw_pytorch).

I haven't written a proper demo yet, but if this is something people would like and would be relatively easy I could try to make time for it.

  • 👍4
mblondel commented
Maghoumi commented

@mblondel First, thanks again for your great paper!

I've been extensively using it for my research. I just publicly released a CUDA implementation of your algorithm for PyTorch that can run up to 100x faster than other available implementations. It's availble here: https://github.com/Maghoumi/pytorch-softdtw-cuda

Hope this helps others!

  • 👍6