SoftDTWLoss Training Calculation

[cvarandia]

Hello. Just wanted to clarify.

Is the SoftDTWLoss function designed to accept arrays of time-series? Since during training and testing in Chainer, the loss is computed per batch (true and predicted), the input to the SoftDTWLoss is a batch of time-series labels and predictions. Does the SoftDTWLoss function consider the whole batch as 1 pair of time series or does it treat each pair of time-series (true and predicted) in a batch pair-wise?

Thank you.

[mblondel]

Hi! As you can see here, I'm looping over the batch in the example code. So SoftDTWLoss is between two time series, not between two batches of time series. I think back then I made this choice because the batch has to be a Python list (since each time series can have a different length).

[cvarandia]

So in order to get the loss of the whole batch, I need to iterate over all the sequences in the batch. And for each iteration, I need to reshape each sequence to (-1,1) first before getting the SoftDTWLoss for each pair of sequences to be compared? Thank you for clarifying.

[mblondel]

SoftDTWLoss expects time series of shape (length, n_dimensions), so if your time series are one dimensional and have shape (length,), you indeed need a reshape.