Latent SDE for irregularly spaced time series? about torchsde HOT 2 CLOSED

Handling irregular time series is generally relatively straightforward. For example if you have some observations x_i \in \mathbb{R}^d made at timestamps t_i \in \mathbb{R}, then you can concatenate (t_i, xi) together and just run an RNN over them. *

torchcde/examples/irregular_data.py is a nice example written in PyTorch, which also handles things like variable-length time series and missing data. (It's written in the context of neural CDEs -- moving on to that now.)

Good choices of encoder would be GRUs, ODE-RNNs, or neural CDEs. (In increasing levels of generality.)

It sounds like you're already familiar with the first two. References for the last one: original NCDE paper, follow up paper, torchcde library.

* footnote: it will often actually be (t_i - t{i-1}, xi), i.e. we concatenate the delta-timestamp rather than the timestamp itself.

from torchsde.

Awesome, thanks for pointing out the references!

from torchsde.