Adds util to compute binary predictive posterior variance

PiperOrigin-RevId: 339414502

edward2/tensorflow/layers/utils.py

@@ -345,6 +345,27 @@ def smart_constant_value(pred):
   return pred_value
 
 
+def mean_field_binary_predictive_variance(logits, covmat, mean_field_factor=1.):
+  """Compute predictive variance for Laplace-approximated logit posterior, assuming sigmoid link.
+
+  Arguments:
+    logits: A float tensor of shape (batch_size, num_classes).
+    covmat: A float tensor of shape (batch_size, batch_size).
+    mean_field_factor: The scale factor for mean-field approximation, used to
+      adjust the influence of posterior variance in posterior mean
+      approximation.
+
+  Returns:
+   Mean-field posterior variance.
+
+  """
+  logits_scale = tf.sqrt(1. + tf.linalg.diag_part(covmat) * mean_field_factor)
+  logits = logits / tf.expand_dims(logits_scale, axis=-1)
+  posterior_mean = tf.sigmoid(tf.squeeze(logits, axis=(1,)))
+
+  return posterior_mean * (1 - posterior_mean) * (1 / logits_scale)
+
+
 def mean_field_logits(logits, covmat, mean_field_factor=1.):
   """Adjust the SNGP logits so its softmax approximates posterior mean [1].
 
@@ -356,7 +377,7 @@ def mean_field_logits(logits, covmat, mean_field_factor=1.):
       approximation.
 
   Returns:
-    True or False if `pred` has a constant boolean value, None otherwise.
+    Calibrated logits.
 
   """
   logits_scale = tf.sqrt(1. + tf.linalg.diag_part(covmat) * mean_field_factor)