sleepwalking / pytorch-softdtw Goto Github PK

I am a greenhand of pytorch, however, I now get x and y of (5089) size. 50 is the length of the array, and 8*9 is the dimension of the feature. How can I operate their dimensions so that I can input the criterion? Thx a lot.

assert len(x.shape) == len(y.shape) squeeze = False if len(x.shape) < 3: x = x.unsqueeze(0) y = y.unsqueeze(0) squeeze = True

Thanks for your contribution, if you can solve my problem I'll be grateful!

Thanks again for your implementation!

Just wanted to let you know that I just released a fast CUDA implementation that's based on your code: https://github.com/Maghoumi/pytorch-softdtw-cuda

I made sure to give you credit and include a link to your repo wherever I included your CPU implementation. If there's any other note in particular that you'd like me to include, please let me know.

I would like to use softdtw for alignment, not as a loss function.

You mention "the derivative is the same as the expected DTW path".

I am new to pytorch, how do I get this derivative? I tried the following,

out = torch.tensor([0.0,0,1,0], requires_grad=True).reshape((1,4,1))
target = torch.tensor([0.0,1,0,0]).reshape((1,4,1))

criterion = SoftDTW(gamma=1.0)
loss = criterion(out, target)
m = loss.backward()
print(m)

which prints "None".

Thank for your source code!

As you said that: "More goodies: check out pytorch-softdtw-cuda by Maghoumi,"

So does it mean this source code is just for CPU computation?