penalty_lbfgs_optimizer.py 文件源码

def update_opt(self, loss, target, leq_constraint, inputs, constraint_name="constraint", *args, **kwargs):
        """
        :param loss: Symbolic expression for the loss function.
        :param target: A parameterized object to optimize over. It should implement methods of the
        :class:`rllab.core.paramerized.Parameterized` class.
        :param leq_constraint: A constraint provided as a tuple (f, epsilon), of the form f(*inputs) <= epsilon.
        :param inputs: A list of symbolic variables as inputs
        :return: No return value.
        """
        constraint_term, constraint_value = leq_constraint
        penalty_var = TT.scalar("penalty")
        penalized_loss = loss + penalty_var * constraint_term

        self._target = target
        self._max_constraint_val = constraint_value
        self._constraint_name = constraint_name

        def get_opt_output():
            flat_grad = flatten_tensor_variables(theano.grad(
                penalized_loss, target.get_params(trainable=True), disconnected_inputs='ignore'
            ))
            return [penalized_loss.astype('float64'), flat_grad.astype('float64')]

        self._opt_fun = lazydict(
            f_loss=lambda: compile_function(inputs, loss, log_name="f_loss"),
            f_constraint=lambda: compile_function(inputs, constraint_term, log_name="f_constraint"),
            f_penalized_loss=lambda: compile_function(
                inputs=inputs + [penalty_var],
                outputs=[penalized_loss, loss, constraint_term],
                log_name="f_penalized_loss",
            ),
            f_opt=lambda: compile_function(
                inputs=inputs + [penalty_var],
                outputs=get_opt_output(),
                log_name="f_opt"
            )
        )