FunctionLogger.batch_call

FunctionLogger.batch_call(x: ndarray, f_vals=None)

Evaluate and cache a batch of points in input order.

Parameters:

xnp.ndarray

A nonempty array with shape (N, D) in transformed coordinates.

f_valsarray-like, optional

Values of the logged function at the corresponding rows of x, before any parameter-transform Jacobian is added. It must have length N; NaN rows are evaluated by the target. At uncertainty handling level 2 every row has to be NaN: a supplied value comes without the SD that an observation needs at that level.

Returns:

valuesnp.ndarray

The recorded values in transformed coordinates, with shape (N,).

SDsnp.ndarray or None

The target SD for each row when noise handling is enabled, otherwise None.

indicesnp.ndarray

Cache indices corresponding to the input rows.

Raises:

RuntimeError

If this logger was not created with vectorized_target=True.

ValueError

If the input or any supplied or returned value has an invalid shape or value, or if f_vals supplies a value at uncertainty handling level 2. The supplied values are validated before the target is called, and its outputs before any row is recorded.

For the logger owned by a VBMC instance, the logged function is VBMC’s assembled log joint, so values passed directly to batch_call include any separately supplied prior. This differs from VBMC(precomputed_evaluations=(X, y)): there, y contains outputs of the log_density argument, and VBMC adds a separately supplied prior once. See the VBMC interface for that higher-level workflow.