tf.contrib.distributions.moving_mean_variance

Compute exponentially weighted moving {mean,variance} of a streaming value.

tf.contrib.distributions.moving_mean_variance(
    value, decay, collections=None, name=None
)

The exponentially-weighting moving mean_var and variance_var are updated by value according to the following recurrence:

variance_var = decay * (variance_var + (1-decay) * (value - mean_var)**2)
mean_var     = decay * mean_var + (1 - decay) * value

Note: mean_var is updated after variance_var, i.e., variance_var uses the lag-1 mean.

For derivation justification, see [Finch (2009; Eq. 143)][1].

Unlike assign_moving_mean_variance, this function handles variable creation.

Args
value	float-like Tensor. Same shape as mean_var and variance_var.
decay	A float-like Tensor. The moving mean decay. Typically close to 1., e.g., 0.999.
collections	Python list of graph-collections keys to which the internal variables mean_var and variance_var are added. Default value is [GraphKeys.GLOBAL_VARIABLES].
name	Optional name of the returned operation.

Returns
mean_var	Variable representing the value-updated exponentially weighted moving mean.
variance_var	Variable representing the value-updated exponentially weighted moving variance.

Raises
TypeError	if value_var does not have float type dtype.
TypeError	if value, decay have different base_dtype.

References

[1]: Tony Finch. Incremental calculation of weighted mean and variance. Technical Report, 2009. http://people.ds.cam.ac.uk/fanf2/hermes/doc/antiforgery/stats.pdf