Method _run

caffe2/python/optimizer.py:1432–1474

(self, net, param_init_net, param_info)

        self.init_kwargs = kwargs

    def _run(self, net, param_init_net, param_info):
        param = param_info.blob
        grad = param_info.grad

        if self.alpha <= 0:
            return

        lr, _ = self.build_lr(
            net,
            param_init_net,
            base_learning_rate=self.alpha,
            policy=self.policy,
            **(self.init_kwargs)
        )

        moment = param_init_net.ConstantFill(
            [param], str(param) + "_squared_moment", value=0.0
        )

        moment_update = param_init_net.ConstantFill(
            [param], str(param) + "_squared_moment_update", value=0.0
        )

        self._aux_params.local.append(moment)
        self._aux_params.local.append(moment_update)

        if isinstance(grad, core.GradientSlice):
            grad = self.dedup(net, self.sparse_dedup_aggregator, grad)
            net.SparseAdadelta(
                [param, moment, moment_update, grad.indices, grad.values, lr],
                [param, moment, moment_update],
                epsilon=self.epsilon,
                decay=self.decay,
                engine=self.engine,
            )
        else:
            net.Adadelta(
                [param, moment, moment_update, grad, lr],
                [param, moment, moment_update],
                epsilon=self.epsilon,
                decay=self.decay,
                engine=self.engine,
            )

    def scale_learning_rate(self, scale):
        self.alpha *= scale