tf.contrib.metrics.streaming_covariance( predictions, labels, weights=None, metrics_collections=None, updates_collections=None, name=None )
Computes the unbiased sample covariance between
streaming_covariance function creates four local variables,
count, which are used to
compute the sample covariance between predictions and labels across multiple
batches of data. The covariance is ultimately returned as an idempotent
operation that simply divides
count - 1. We use
count - 1
in order to get an unbiased estimate.
The algorithm used for this online computation is described in
Specifically, the formula used to combine two sample comoments is
C_AB = C_A + C_B + (E[x_A] - E[x_B]) * (E[y_A] - E[y_B]) * n_A * n_B / n_AB
The comoment for a single batch of data is simply
sum((x - E[x]) * (y - E[y])), optionally weighted.
weights is not None, then it is used to compute weighted comoments,
means, and count. NOTE: these weights are treated as "frequency weights", as
opposed to "reliability weights". See discussion of the difference on
To facilitate the computation of covariance across multiple batches of data,
the function creates an
update_op operation, which updates underlying
variables and returns the updated covariance.
Tensorof arbitrary size.
Tensorof the same size as
Tensorindicating the frequency with which an example is sampled. Rank must be 0, or the same rank as
labels, and must be broadcastable to
labels(i.e., all dimensions must be either
1, or the same as the corresponding
metrics_collections: An optional list of collections that the metric value variable should be added to.
updates_collections: An optional list of collections that the metric update ops should be added to.
name: An optional variable_scope name.
Tensorrepresenting the current unbiased sample covariance,
update_op: An operation that updates the local variables appropriately.
ValueError: If labels and predictions are of different sizes or if either
updates_collectionsare not a list or tuple.