提交 11b77a81 编写于 作者: L liuzhidan

update the notation of Adaclip

上级 a2a1b56b
...@@ -50,7 +50,7 @@ class ClipMechanismsFactory: ...@@ -50,7 +50,7 @@ class ClipMechanismsFactory:
learning_rate(float): Learning rate of update norm clip. Default: 0.001. learning_rate(float): Learning rate of update norm clip. Default: 0.001.
target_unclipped_quantile(float): Target quantile of norm clip. Default: 0.9. target_unclipped_quantile(float): Target quantile of norm clip. Default: 0.9.
fraction_stddev(float): The stddev of Gaussian normal which used in fraction_stddev(float): The stddev of Gaussian normal which used in
empirical_fraction, the formula is $empirical_fraction + N(0, fraction_stddev)$. empirical_fraction, the formula is :math:`empirical fraction + N(0, fraction sstddev)`.
Default: 0.01. Default: 0.01.
seed(int): Original random seed, if seed=0 random normal will use secure seed(int): Original random seed, if seed=0 random normal will use secure
random number. IF seed!=0 random normal will generate values using random number. IF seed!=0 random normal will generate values using
...@@ -342,10 +342,10 @@ class _MechanismsParamsUpdater(Cell): ...@@ -342,10 +342,10 @@ class _MechanismsParamsUpdater(Cell):
class AdaClippingWithGaussianRandom(Cell): class AdaClippingWithGaussianRandom(Cell):
""" """
Adaptive clipping. If `decay_policy` is 'Linear', the update formula is Adaptive clipping. If `decay_policy` is 'Linear', the update formula :math:`norm bound = norm bound -
norm_bound = norm_bound - learning_rate*(beta - target_unclipped_quantile). learning rate*(beta - target unclipped quantile)`.
If `decay_policy` is 'Geometric', the update formula is norm_bound = If `decay_policy` is 'Geometric', the update formula is :math:`norm bound =
norm_bound*exp(-learning_rate*(empirical_fraction - target_unclipped_quantile)). norm bound*exp(-learning rate*(empirical fraction - target unclipped quantile))`.
where beta is the empirical fraction of samples with the value at most where beta is the empirical fraction of samples with the value at most
`target_unclipped_quantile`. `target_unclipped_quantile`.
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册