model.py 文件源码

def build_bprop_graph(self):
        optimizer = self.get_optimizer()
        # there are either costs assigned to specific params
        # OR let blocks do the gradient
        costs = self.link_here('costs').keys()

        isinstance_check = [isinstance(c, ParametersLink) for c in costs]
        if any(isinstance_check):
            assert all(isinstance_check), "Some costs have parameters associated "+\
                    "to them and others don't. All costs need to be binded."
            grads = OrderedDict()
            for cost in costs:
                grads.update(zip(cost.parameters,
                                 theano.grad(cost.model_var, cost.params)))
            cost = None
        else:
            cost = sum(costs)
            grads = None

        algorithm = GradientDescent(
            cost=cost, gradients=grads,
            parameters=self.model_parameters,
            step_rule=optimizer)

        self.algorithm = algorithm