Join us at TensorFlow World, Oct 28-31. Use code TF20 for 20% off select passes. Register now

tfp.math.ode.Solver

Class Solver

Base class for an ODE solver.

Defined in python/math/ode/base.py.

Methods

solve

solve(
    ode_fn,
    initial_time,
    initial_state,
    solution_times,
    jacobian_fn=None,
    jacobian_sparsity=None,
    batch_ndims=None,
    previous_solver_internal_state=None
)

Solves an initial value problem.

An initial value problem consists of a system of ODEs and an initial condition:

dy/dt(t) = ode_fn(t, y(t))
y(initial_time) = initial_state

Here, t (also called time) is a scalar float Tensor and y(t) (also called the state at time t) is an N-D float or complex Tensor.

Args:

  • ode_fn: Function of the form ode_fn(t, y). The input t is a scalar float Tensor. The input y and output are both Tensors with the same shape and dtype as initial_state.
  • initial_time: Scalar float Tensor specifying the initial time.
  • initial_state: N-D float or complex Tensor specifying the initial state. The dtype of initial_state must be complex for problems with complex-valued states (even if the initial state is real).
  • solution_times: 1-D float Tensor specifying a list of times. The solver stores the computed state at each of these times in the returned Results object. Must satisfy initial_time <= solution_times[0] and solution_times[i] < solution_times[i+1]. Alternatively, the user can pass tfp.math.ode.ChosenBySolver(final_time) where final_time is a scalar float Tensor satisfying initial_time < final_time. Doing so requests that the solver automatically choose suitable times up to and including final_time at which to store the computed state.
  • jacobian_fn: Optional function of the form jacobian_fn(t, y). The input t is a scalar float Tensor. The input y has the same shape and dtype as initial_state. The output is a 2N-D Tensor whose shape is initial_state.shape + initial_state.shape and whose dtype is the same as initial_state. In particular, the (i1, ..., iN, j1, ..., jN)-th entry of jacobian_fn(t, y) is the derivative of the (i1, ..., iN)-th entry of ode_fn(t, y) with respect to the (j1, ..., jN)-th entry of y. If this argument is left unspecified, the solver automatically computes the Jacobian if and when it is needed. Default value: None.
  • jacobian_sparsity: Optional 2N-D boolean Tensor whose shape is initial_state.shape + initial_state.shape specifying the sparsity pattern of the Jacobian. This argument is ignored if jacobian_fn is specified. Default value: None.
  • batch_ndims: Optional nonnegative integer. When specified, the first batch_ndims dimensions of initial_state are batch dimensions. Default value: None.
  • previous_solver_internal_state: Optional solver-specific argument used to warm-start this invocation of solve. Default value: None.

Returns:

Object of type Results.