torchdyn

Adds torch 2.0 compatibility

Deprecate python 3.7

Incorporating #149 and #150 fixing bug preventing save/load of ODEProblem objects ( raised in issue #122).

• Added the ability to explicitly pass parameters in the form of a dictionary at each ODE solver step
• fixed_odeint can now accept dictionaries as state
• added save_at argument to odeint to have the solver save the solution only at specified time instants
• Data control variable u in NeuralODE has been changed to _control
• Added support for Python 3.7
• various bug fixes

torchdyn 1.0 is finally out!

Minor fixes. More prep for 1.0. Updates PyPI installable version of torchdyn.

More tutorials. MultipleShootingProblem is in. Several fixes to the numerics.

Edit with latest fixes.

The current release greatly extends the functionality offered by torchdyn, planning for a larger 1.0 release later this month (with multiple shooting layers, root finding, rudimentary implicit solvers, hybrid neural DE solvers).

• New ODEProblem for standard IVPs. Alternative sensitivity algorithms rely ODEProblems for sanity, and are specified at __init__. Backpropagating through the functional API (direct calls to odeint) only uses reverse-mode autodiff.
• A new interpolated adjoint variant relying on torchcde utilities.
• New functional odeint API with adaptive-step solvers: dopri5, tsit5.
• New odeint_hybrid for multi-mode hybrid ODEs with stochastic or deterministic events.
• Seamless hypersolver integration with odeint API, with same call signature!

Some more minor or less interesting changes:

• Streamlining of tests and new tutorial automated notebook validation.
• s changed to t for more clarity and compatibility with the current literature.

Expect bugs and rough edges (and some WIP hidden in the library). The 1.0 will clean those up.