Adjoint methods

[ToennisStef]

Hello everyone,
I have a question regarding the differentiable integrators. For the backwards pass through the time integration in the ODE solvers, are adjoint methods implemented to get efficient sensitivities, for example for the "paramter estimation problem of an ode system" example?

Otherwise thanks to the creators of this beautiful package! I am considering to fully switch to Neuromancer and use this in addition with pyro to have a probablistic modeling framework for stochastic differential equations. I was previously leaning into using casadi + pytorch.autograd.function + pyro, but this seems like the more natural choice, especially if i would have the direct cuda capabilities here.
So thanks again!

[drgona]

@ToennisStef, yes, Neuromancer supports both the adjoint method and backpropagation through time differentiation of the ODE
integrators.

See the implementation details here

and tutorial examples here