search

tslearn-team / tslearn

The machine learning toolkit for time series analysis in Python
https://tslearn.readthedocs.io
BSD 2-Clause "Simplified" License
2.83k stars 331 forks source link

Compute SoftDTWLossPyTorch with normalization option and time series of different lengths #473

Open denisbeslic opened 11 months ago

denisbeslic commented 11 months ago

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):

Additional context

YannCabanes commented 9 months ago

Hello @denisbeslic, thanks a lot for reporting this bug! I am dealing with this issue in PR https://github.com/tslearn-team/tslearn/pull/478.