Skip to content

dq.gradient.CheckpointAutograd 

CheckpointAutograd(ncheckpoints: int | None = None)

Checkpointed automatic differentiation.

With this option, the gradient is computed by automatically differentiating through the internals of the solver. The difference with the standard automatic differentiation (see dq.gradient.Autograd) is that a checkpointing scheme is used to reduce the memory usage of the backpropagation.

Note

For most problems this is the preferred technique for backpropagating through the quantum solvers.

Warning

This cannot be forward-mode autodifferentiated (e.g. using jax.jvp ). Try using dq.gradient.Autograd if that is something you need.

Note

For Diffrax-based solvers, this falls back to the diffrax.RecursiveCheckpointAdjoint option.

Parameters

  • ncheckpoints –

    Number of checkpoints to use. The amount of memory used by the differential equation solve will be roughly equal to the number of checkpoints multiplied by the size of the state. You can speed up backpropagation by allocating more checkpoints, so it makes sense to set as many checkpoints as you have memory for. This value is set to None by default, in which case it will be set to log(max_steps), for which a theoretical result is available guaranteeing that backpropagation will take O(n_steps log(n_steps)) time in the number of steps n_steps <= max_steps.