Warn if taking softmax over broadcastable dimension

theano/tensor/nnet/nnet.py

@@ -802,6 +802,9 @@ def softmax_graph(c):
 
 
 def softmax(c):
+    c = as_tensor_variable(c)
+    if c.broadcastable[-1]:
+        warnings.warn("The softmax is appllied on a dimension of shape 1, which doesn't have a semantic meaning.")
     return softmax_op(c)