with in-place option

python/paddle/v2/framework/tests/gradient_checker.py

@@ -32,7 +32,8 @@ def get_numeric_gradient(op,
                          output_name,
                          input_to_check,
                          delta=0.005,
-                         local_scope=None):
+                         local_scope=None,
+                         in_place=False):
     """
     Get Numeric Gradient for an operator's input.
     
@@ -90,6 +91,7 @@ def get_numeric_gradient(op,
     # we only compute gradient of one element each time.
     # we use a for loop to compute the gradient of every element.
     for i in xrange(tensor_size):
+        if in_place:
             for var_name in input_values:
                 tensor_ = local_scope.find_var(var_name).get_tensor()
                 tensor_.set(numpy.copy(input_values[var_name]), core.CPUPlace())
@@ -102,6 +104,7 @@ def get_numeric_gradient(op,
         y_pos = get_output()
 
         # plus delta to this element, run op and get the sum of the result tensor.
+        if in_place:
             for var_name in input_values:
                 tensor_ = local_scope.find_var(var_name).get_tensor()
                 tensor_.set(numpy.copy(input_values[var_name]), core.CPUPlace())
@@ -257,6 +260,7 @@ class GradientChecker(unittest.TestCase):
                    output_name,
                    no_grad_set=None,
                    only_cpu=False,
+                   in_place=False,
                    max_relative_error=0.005):
         """
         :param forward_op: used to create backward_op
@@ -289,7 +293,8 @@ class GradientChecker(unittest.TestCase):
 
         # get numerical gradients
         numeric_grads = [
-            get_numeric_gradient(forward_op, input_vars, output_name, name)
+            get_numeric_gradient(
+                forward_op, input_vars, output_name, name, in_place=in_place)
             for name in inputs_to_check
         ]
 

python/paddle/v2/framework/tests/test_scatter_op.py

@@ -31,7 +31,8 @@ class TestScatterGradOp(GradientChecker):
         output_np[index_np] += updates_np
         inputs = {'Ref': ref_np, 'Index': index_np, 'Updates': updates_np}
         # check gradient
-        self.check_grad(op, inputs, set(["Updates", "Ref"]), "Out")
+        self.check_grad(
+            op, inputs, set(["Updates", "Ref"]), "Out", in_place=True)
 
 
 if __name__ == "__main__":