Class GFtrlOptimizer

caffe2/python/optimizer.py:1537–1580

Group Lasso FTRL Optimizer.

class GFtrlOptimizer(Optimizer):
    """Group Lasso FTRL Optimizer."""

    def __init__(
        self,
        alpha=0.01,
        beta=1e-4,
        lambda1=0,
        lambda2=0,
        sparse_dedup_aggregator=None,
        engine="",
    ):
        super().__init__()
        self.alpha = alpha
        self.beta = beta
        self.lambda1 = lambda1
        self.lambda2 = lambda2
        self.sparse_dedup_aggregator = sparse_dedup_aggregator
        self.engine = engine

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

        if self.alpha <= 0:
            return

        nz = param_init_net.ConstantFill(
            [param], str(param) + "_gftrl_nz", extra_shape=[2], value=0.0
        )
        self._aux_params.local.append(nz)
        net.GFtrl(
            [param, nz, grad],
            [param, nz],
            engine=self.engine,
            alpha=self.alpha,
            beta=self.beta,
            lambda1=self.lambda1,
            lambda2=self.lambda2,
        )

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


class AdamOptimizer(Optimizer):