first gradient unittest

contrib/Python/cntk/context.py

@@ -285,14 +285,14 @@ class AbstractContext(object, metaclass=ABCMeta):
         pass
 
     @abstractmethod
-    def eval(self, node, reader=None, node_unit_test=False):
+    def eval(self, node, reader=None, backward_pass=False):
         '''
         Abstract method for the action write. It evaluated the passed node on the
         data provided by the reader.
         :param node: the node to evaluate.
         :param reader: the reader to use for this action. Alternatively, you
         can attach a reader directly to the input node.
-        :param node_unit_test: set to True if you want to output the gradient of a node (backward pass)
+        :param backward_pass: set to True if you want to output the gradient of a node (backward pass)
         Returns the output generated by `node`
         '''
         pass
@@ -543,7 +543,7 @@ class Context(AbstractContext):
 
         return expected_shape, expected_size
 
-    def eval(self, node, reader=None, node_unit_test=False, input_name=None):
+    def eval(self, node, reader=None, backward_pass=False, input_name=None):
         '''
         Run the write action locally to evaluate the passed node and returning
         the data it produced.
@@ -551,8 +551,8 @@ class Context(AbstractContext):
         :param node: the node to evaluate.
         :param reader: the reader to use for this action. Alternatively, you
         can attach a reader directly to the input node.
-        :param node_unit_test: set to True if you want to output the gradient of a node (backward pass)        
+        :param backward_pass: set to True if you want to output the gradient of a node (backward pass)
-        :input_name: if node_unit_test is True then input_node should contain the input name that
+        :input_name: if backward_pass is True then input_node should contain the input name that
         the gradient is performed with respect to.
         Returns the output generated by `node`
         '''
@@ -564,15 +564,15 @@ class Context(AbstractContext):
         orig_node_tag = node.tag if hasattr(node, 'tag') else None
         node.tag = 'output'
 
-        config_content = self._generate_eval_config(node, reader, node_unit_test)
+        config_content = self._generate_eval_config(node, reader, backward_pass)
-        output = self._call_cntk(CNTK_EVAL_CONFIG_FILENAME, config_content)
+        self._call_cntk(CNTK_EVAL_CONFIG_FILENAME, config_content)
 
         node.tag = orig_node_tag
 
         n = input_name.var_name if isinstance(input_name, ComputationNode) else input_name
         out_name = os.path.join(
             self.directory, CNTK_OUTPUT_FILENAME + '.' + \
-                ((n + '.grad') if node_unit_test  else node.var_name))        
+                ((n + '.grad') if backward_pass  else node.var_name))        
 
         result_content = open(out_name).read()
         data = Context._parse_result_output(result_content)

contrib/Python/cntk/test/test_ops.py

@@ -11,14 +11,17 @@ C = constant
 I = input
 AA = np.asarray
 
-def _test(root_node, expected, clean_up=True):
+def _test(root_node, expected, clean_up=True, backward_pass = False, input_node = None):
     with get_new_context() as ctx:
         ctx.clean_up = clean_up
         assert not ctx.input_nodes
-        result = ctx.eval(root_node)
+        result = ctx.eval(root_node, None, backward_pass, input_node)
 
         assert len(result) == len(expected)
-        for res, exp in zip(result, expected):
+        for res, exp in zip(result, expected):            
+            print(res)
+            print(exp)
+            print("===========#####")
             assert np.allclose(res, exp)
             assert res.shape == AA(exp).shape
 
@@ -84,9 +87,17 @@ def test_op_add_input_constant(left_arg, right_arg):
      ],  
       2), 
     ])
-
 def test_op_mul_input_seq(left_arg, right_arg):
     expected = [AA(elem)*right_arg for elem in left_arg]
     result = I(left_arg, has_sequence_dimension=True) * right_arg
     _test(result, expected, False)
 
+@pytest.mark.parametrize("left_arg, right_arg", [
+    ([0],[0]),              # grad(Cos(0)) = -Sin(0) = 0
+    ([1.57079633],[-1]),     # grad(Cos(pi/2)) = -Sin(pi/2) = -1
+    ([3.14159265],[0]),    # grad(Cos(pi)) = -Sin(pi) = 0 
+    ])
+def test_cosine_backward(left_arg, right_arg):
+     i = I([left_arg], has_sequence_dimension=False)
+     n = Cosine(i)
+     _test(n, [[AA(right_arg)]], clean_up=False, backward_pass = True, input_node = i)