define mean for CategoricalPd (as softmax of logits)

baselines/common/distributions.py

@@ -146,6 +146,10 @@ class CategoricalPd(Pd):
         return self.logits
     def mode(self):
         return tf.argmax(self.logits, axis=-1)
+
+    @property
+    def mean(self):
+        return tf.nn.softmax(self.logits)
     def neglogp(self, x):
         # return tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.logits, labels=x)
         # Note: we can't use sparse_softmax_cross_entropy_with_logits because