def test_adding_with_same_node(self):
# Forward
op = Add(Add(self.a, self.a), Add(self.a, self.a))
op.register(op)
op.new_context()
expected = self.a.data * 4
actual = op.forward()
self.assertTrue((expected == actual).all())
# Backward
gradient = np.ones((2, 3))
op.accumulate(op, gradient)
expected = np.full((2, 3), 4)
self.assertTrue((expected == self.a.gradient).all())
def test_adding_same_shape(self):
# Forward
op = Add(self.a, self.b)
op.register(op)
op.new_context()
expected = np.array([[2, 3, 7], [4, 5, 6]])
actual = op.forward()
self.assertTrue((expected == actual).all())
# Backward
gradient = np.ones((2, 3))
op.accumulate(op, gradient)
actual1 = self.a.gradient
actual2 = self.b.gradient
self.assertTrue((gradient == actual1).all())
self.assertTrue((gradient == actual2).all())
def test_adding_matrix_with_different_number_of_rows(self):
# Forward
op = Add(self.a, self.c)
op.register(op)
op.new_context()
expected = np.array([[2, 3, 7], [4, 5, 6]])
actual = op.forward()
self.assertTrue((expected == actual).all())
# Backward
gradient = np.ones((2, 3))
op.accumulate(op, gradient)
actual1 = self.a.gradient
expected2 = np.full((1, 3), 2)
actual2 = self.c.gradient
self.assertTrue((gradient == actual1).all())
self.assertTrue((expected2 == actual2).all())
def test_adding_nested_operations(self):
# Forward
op = Add(self.a, Add(self.b, self.c))
op.register(op)
op.new_context()
expected = np.array([[3, 4, 8], [5, 6, 7]])
actual = op.forward()
self.assertTrue((expected == actual).all())
# Backward
gradient = np.ones((2, 3))
op.accumulate(op, gradient)
expected1 = gradient
expected2 = gradient
expected3 = np.full((1, 3), 2)
self.assertTrue((expected1 == self.a.gradient).all())
self.assertTrue((expected2 == self.b.gradient).all())
self.assertTrue((expected3 == self.c.gradient).all())
def test_adding_matrix_with_scalar(self):
# Forward
op = Add(self.a, self.e)
op.register(op)
op.new_context()
expected = np.array([[3, 4, 8], [5, 6, 7]])
actual = op.forward()
self.assertTrue((expected == actual).all())
# Backward
gradient = np.ones((2, 3))
op.accumulate(op, gradient)
expected1 = gradient
actual1 = self.a.gradient
expected2 = np.array(6)
actual2 = self.e.gradient
self.assertTrue((expected1 == actual1).all())
self.assertTrue((expected2 == actual2).all())
def test_adding_nested_operations_with_repeated_nodes(self):
# Forward
subexp1 = Add(self.a, self.b)
subexp2 = Add(self.b, self.c)
op = Add(subexp1, subexp2)
op.register(op)
op.new_context()
expected = np.array([[4, 5, 9], [6, 7, 8]])
actual = op.forward()
self.assertTrue((expected == actual).all())
# Backward
gradient = np.ones((2, 3))
op.accumulate(op, gradient)
expected1 = gradient
expected2 = np.full((2, 3), 2)
expected3 = np.full((1, 3), 2)
self.assertTrue((expected1 == self.a.gradient).all())
self.assertTrue((expected2 == self.b.gradient).all())
self.assertTrue((expected3 == self.c.gradient).all())