Created by: lcy-seso

There is a code snippet from paddle/trainer_config_helpers/attrs.py https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/trainer_config_helpers/attrs.py#L139:

    def __init__(self,
                 name=None,
                 is_static=False,
                 initial_std=None,
                 initial_mean=None,
                 initial_max=None,
                 initial_min=None,
                 l1_rate=None,
                 l2_rate=None,
                 learning_rate=None,
                 momentum=None,
                 gradient_clipping_threshold=None,
                 sparse_update=False,
                 update_hooks=None,
                 initializer=None):
        self.attr = {}

        if is_static:
            self.attr['is_static'] = True

        if initial_std is None and initial_mean is None and initial_max \
                is None and initial_min is None:
            self.attr['initial_smart'] = True

• in training, I may prefer another initialization strategy, such as initialization from a Gaussian distribution with a default mean and std, rather than use smart initialization.

• to do this, I will set the global initialization std and mean.

• The implementation above ignores the globally set mean and std. If I just use paddle.attr.Param to tune the learning rate or any other parameters except initial_std, initial_mean, initial_max, and initial_min like below. The codes then force me to use smart initialization. It ignores the default initialization std I set.

  param_attr=paddle.attr.Param(learning_rate=0.1)