def test_grad_on_X(self):
self.layer.forward(self.X)
d_X = self.layer.backward(self.grad_top)
assert_array_almost_equal(
numerical_grad(self.layer.forward, self.X),
d_X
)
def test_backward(self):
self.layer.forward(self.X, self.y)
d_X = self.layer.backward()
layer = self.layer
def forward_as_func_of_X(X):
return layer.forward(X, self.y)
assert_array_almost_equal(numerical_grad(forward_as_func_of_X, self.X),
d_X)
def test_grad_on_b(self):
self.layer.forward(self.X)
self.layer.backward(self.grad_top)
d_b = self.layer.d_b
layer = self.layer
def forward_as_func_of_b(b_):
layer.b = b_
return layer.forward(self.X)
assert_array_almost_equal(
numerical_grad(forward_as_func_of_b, self.b),
d_b
)
def test_grad_on_W(self):
self.layer.forward(self.X)
self.layer.backward(self.grad_top)
d_W = self.layer.d_W
layer = self.layer
def forward_as_func_of_W(W_):
layer.W = W_
return layer.forward(self.X)
assert_array_almost_equal(
numerical_grad(forward_as_func_of_W, self.W),
d_W
)
def test_numerical_grad_ndarray(self):
X = np.array([
[-5, 1, -1, 10, -2],
[8, 10, -12, 3, 1],
[0, 0, 2, -1, 5]], dtype=float)
def dot_Y(X_):
return np.dot(X_, self.Y)
expected_grad_X = np.array([
[6, -5, 8, 11, 11],
[6, -5, 8, 11, 11],
[6, -5, 8, 11, 11]], dtype=float)
assert_array_almost_equal(
numerical_grad(dot_Y, X),
expected_grad_X
)
def test_numerical_grad_scalar(self):
def times_5(x_):
return x_ * 5
self.assertAlmostEqual(numerical_grad(times_5, 12), 5)