More than an issue, this is a question/request. In my lab, we are working with physical systems for which the forward pass can be computed in the physical system, but the backward pass must be computed in a computer that has an approximate model of the physical system (and hence can be back-propagated).
For testing this for small physical system sizes, I was imagining that one would use a full simulation of the system to go forward with no gradient by calling in on odeint on func_physical, a function that represents the ODEs of the physical system with no approximations. This would tell us the trajectory as dictated by the actual physical system. I want to then feed that trajectory to go backwards via the adjoint_method. However, I would like to have the adjoint_method call odeint with a new func_model, a function that represents the APPROXIMATE ODEs of the physical system.
I was wondering if there's an easy way to do this using torchdiffeq? Again, this would be very useful for the physical ML community that may want to use Neural ODEs.
More than an issue, this is a question/request. In my lab, we are working with physical systems for which the forward pass can be computed in the physical system, but the backward pass must be computed in a computer that has an approximate model of the physical system (and hence can be back-propagated).
For testing this for small physical system sizes, I was imagining that one would use a full simulation of the system to go forward with no gradient by calling in on odeint on func_physical, a function that represents the ODEs of the physical system with no approximations. This would tell us the trajectory as dictated by the actual physical system. I want to then feed that trajectory to go backwards via the adjoint_method. However, I would like to have the adjoint_method call odeint with a new func_model, a function that represents the APPROXIMATE ODEs of the physical system.
I was wondering if there's an easy way to do this using torchdiffeq? Again, this would be very useful for the physical ML community that may want to use Neural ODEs.
Thanks!