|View source on GitHub|
Namedtuple of quantities that may be traced from
tfp.math.MinimizeTraceableQuantities( step, loss, gradients, parameters, has_converged, convergence_criterion_state )
These are (in order):
Tensorindex (starting from zero) of the current optimization step.
Tensorvalue returned from the user-provided
gradients: list of
losswith respect to the parameters.
parameters: list of
Tensorvalues of parameters being optimized. This corresponds to the
trainable_variablespassed in to
Tensorof the same shape as
Truevalues corresponding to loss entries that have converged according to the user-provided convergence criterion. If no convergence criterion was specified, this is
convergence_criterion_state: structure of
Tensors containing any auxiliary state (e.g., moving averages of loss or other quantities) maintained by the user-provided convergence criterion.