copy only those inputs marked for destruction

theano/gradient.py

@@ -544,18 +544,22 @@ def _populate_grad_dict(var_to_node_to_idx,\
 
             inputs = node.inputs
 
-            def try_to_copy(var):
-                if hasattr(var,'copy'):
+            # Each Op's grad function requires inputs and output_grads
+            # If the Op destroys any input, but the grad expression uses it,
+            # then chances are the resulting graph will have a dependency
+            # cycle. We avoid this cycle by passing (symbolic) copies of
+            # each destroyed input.
+            try:
+                dinputs = [node.inputs[x[0]] for x in node.op.destroy_map.values()]
+            except AttributeError:
+                dinputs = []
+
+            def try_to_copy_if_needed(var):
+                if var in dinputs and hasattr(var,'copy'):
                     return var.copy()
                 return var
 
-            #inplace ops often have inplace in their expression for the gradient
-            #this can result in cyclical dependencies, ie there not being an order
-            #in which we can run all the resulting inplace ops without destroying
-            #some op's input before the time that it is needed
-            #to get around this, we try to symbolically copy all of the inputs
-            #so it is only the copy that is destroyed
-            inputs = [try_to_copy(ipt) for ipt in inputs ]
+            inputs = [try_to_copy_if_needed(ipt) for ipt in inputs ]
 
             output_grads = [ access_grad_cache(var) for var in node.outputs ]