Patrick Kidger

Yeah, adding our own manual print statements might be one approach. Not super elegant of course so if we did this I'd probably add a global toggle on whether to...

Just chipping into say that a few other (mutually independent) approaches for min/max also come to mind: 1. Define an event that detects when (`dx(t)/dt=`)`f(t, x(t)) = 0`. Plus some...

Nope, just natural cubic splines.

I'm afraid not -- this package is only lightly maintained so realistically I'm not going to be adding new features. I'd suggest developing your own package if this is something...

Good catch. I'd be happy to accept a PR fixing this. Yep, none of this code is very efficient on GPU. My own use case for this was interpolating data...

Right. If you need it to run online then the implementation in this repository won't be suitable, I'm afraid, as natural cubic splines don't satisfy that property: the "future" affects...

That, is you want to interpolate along e.g. the height dimension? Sure: ``` data = torch.rand(batch, channel, height, width) data_perm = data.transpose(1, 3) # shape (batch, width, height, channel) ```...

Along the channel dimension? Sounds a little odd, but sure. `natural_cubic_spline_coeffs` takes in tensors of shape `(..., X, Y)`, corresponding to a (batch of) sequences of length `X`, with each...

More or less. The final call to spline.derivative will be evaluated at all the t points, and those were also specified as being the times of the knots in natural_cubic_spline_coeffs,...

So I frequently use this on relatively long sequences. Linear scaling in time is good for peace of mind, and linear scaling in memory might be a necessity to have...