Michael Poli

Sorry for the late reply, lot of deadlines to take care of. Pinging @Juju-botu as he's also working on hypersolvers currently and is interested in contributing. I'll reach out via...

I agree that this is unnecessary (and adds a small overhead). What would be a use case for this? Logging step times and other metrics via a proper Solution class...

this calls for a `verbose` option perhaps.

Hey, thanks for using `torchdyn`. What you ask is not currently supported with the `DataControl` layer, but can be done pretty easily in two ways: * You can have your...

Glad to hear it worked. Depending on your implementation, it should work just fine and get gradients. It is certainly possible to build PointFlows with `torchdyn` :)

Providing an update here based on our private chat, in case someone else is interested: this should be done using IFT at the fixed-point, implementing a custom backward for the...

Great news @zou3519, we will begin integrating `functorch` in several `torchdyn` classes.

Hi @shekshaa! Normally `odeint` calls return `t_eval, sol`, where `t_eval` is of the same length as your desired evaluation `t_span`. However, you can do the following: ``` f = lambda...

This error is happening while solving the adjoint dynamics for your net. The key lines are [47 onwards](https://github.com/DiffEqML/torchdyn/blob/1ee57d7b6d9b9991234c5b96780c01002f1ea712/torchdyn/numerics/sensitivity.py#L47) ``` xT, λT, μT = sol[-1], grad_output[-1][-1], torch.zeros_like(vf_params) ``` which are then...

Identifying what the difference `1206497 - 1203142 = 3355` represents in terms of _elements_ is key here. The shape `1206497` is determined during initialization of the adjoint as a concat...