Skip to main content

Optimizer

mini_ode::optimizers

Trait Optimizer 

Source 
pub trait Optimizer:
    Send
    + Sync
    + Display {
    // Required method
    fn optimize(
        &self,
        function: &dyn Fn(&Tensor) -> Tensor,
        x0: &Tensor,
    ) -> Result<Tensor>;
}
Expand description

Core trait defining the interface for all optimization algorithms.

This trait abstracts over different optimization strategies (Newton, BFGS, CG, Halley), allowing them to be used interchangeably with implicit ODE solvers. The trait is designed for thread-safe usage with Send + Sync bounds.

§Function Signature

The optimize method accepts:

  • function: A closure taking a 1D tensor x and returning a scalar tensor representing the objective value to minimize. The function should use torch operations compatible with autodifferentiation.
  • x0: Initial guess as a 1D tensor, which determines both the problem dimension and computation device (CPU/GPU).

§Return Value

Returns Ok(optimal_x) containing the optimized parameter vector, or Err(e) if optimization fails due to convergence issues, numerical instability, or invalid inputs.

§Implementation Notes

  • The function is evaluated multiple times during optimization
  • Gradients/Hessians are computed automatically via torch.autograd
  • Tolerance parameters control convergence criteria

Required Methods§

Source

fn optimize( &self, function: &dyn Fn(&Tensor) -> Tensor, x0: &Tensor, ) -> Result<Tensor>

Minimizes an objective function starting from an initial guess.

§Arguments
  • function - A closure that takes a 1D tensor x and returns a scalar tensor representing the objective value to minimize. Must support automatic differentiation.
  • x0 - Initial guess, 1D tensor accepted by function. Determines computation device and floating-point precision.
§Returns

Optimal x that minimizes function, or error if optimization fails.

§Errors

Returns an error if:

  • Input validation fails (non-scalar output, wrong rank, non-finite values)
  • Hessian allocation fails due to insufficient memory
  • Line search fails
  • Maximum iterations reached without convergence
§Panics

May panic if libtorch tensor operations fail unexpectedly (should be rare).

Dyn Compatibility§

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors§