Compute SoftDTWLossPyTorch with normalization option and time series of different lengths

[denisbeslic]

Describe the bug
Hello, I want to use the SoftDTWLoss with normalization (Soft-DTW divergence) for comparing predicted time series of different length. It seems to throw this error.

  File ".../lib/python3.10/site-packages/tslearn/metrics/soft_dtw_loss_pytorch.py", line 146, in forward
    xxy = torch.cat([x, x, y])

RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 99 but got size 100 for tensor number 2 in the list.

Is it possible to generate a SoftDTWLoss for time series with unequal lengths or is this a bug?

To Reproduce

targets = torch.tensor(np.random.randn(16, 100, 1))
prediction = torch.tensor(np.random.randn(16, 99, 1))
loss_func = SoftDTWLossPyTorch(gamma=1.0, normalize=True, dist_func=None)
loss = loss_func(prediction, targets)
print(loss.shape)
print(loss)

Environment (please complete the following information):

• OS: Ubuntu 20.04.5 LTS
• tslearn version 0.6.1

Additional context

[YannCabanes]

Hello @denisbeslic, thanks a lot for reporting this bug!
I am dealing with this issue in PR #478.