fixed bug where taking the gradient of an expression with inplace ops

theano/gradient.py

@@ -543,6 +543,20 @@ def _populate_grad_dict(var_to_node_to_idx,\
         if node not in term_dict:
 
             inputs = node.inputs
+
+            def try_to_copy(var):
+                if 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 ]
+
             output_grads = [ access_grad_cache(var) for var in node.outputs ]
             input_grads = node.op.grad(inputs, output_grads)
 
@@ -1057,6 +1071,7 @@ def verify_grad(fun, pt, n_tests=2, rng=None, eps=None,
     symbolic_grad = grad(cost, tensor_pt, g_cost,
                          disconnected_inputs='ignore')
 
+
     grad_fn = function(tensor_pt, symbolic_grad)
 
     for test_num in xrange(n_tests):