def test_output_error_passed_back_in_the_chain_matches_input_dim(self):
x, err, lr = np.array([1, 2, 3]), np.array([
0.5,
]), 3
layer = Output(x)
layer.solve_bwd(err, lr)
self.assertEqual(layer.e.shape, layer.x.shape)
def test_output_gradient_descent(self):
x, err, lr = np.array([1, 2, 3]), np.array([
0.5,
]), 3
layer = Output(x)
layer.solve_bwd(err, lr)
expected = -lr * np.repeat(err, x.shape[0]) * layer.partial_s * layer.x
self.assertTrue((expected == layer.delta_w).all())
def test_output_acc_error_to_be_passed_back_in_the_chain(self):
x, err, lr = np.array([1, 2, 3]), np.array([
0.5,
]), 3
layer = Output(x)
layer.solve_bwd(err, lr)
expected = np.repeat(err, x.shape[0]) * layer.partial_s * layer.w
self.assertTrue((expected == layer.e).all())
def test_output_partial_s(self):
x = np.array([1, 2, 3])
layer = Output(x)
layer.solve_bwd(np.array([
0.5,
]))
partial_s = layer.s * (1 - layer.s)
self.assertTrue((layer.partial_s == partial_s).all())
self.assertEqual(layer.partial_s.shape, x.shape)
def test_output_solve_bwd_check_args_shape(self):
layer = Output(np.array([1, 2]))
regex = 'The shape for net error should be 1'
with self.assertRaisesRegex(ValueError, regex):
layer.solve_bwd(np.array([1, 2]))
def test_output_solve_bwd_check_args_np(self):
layer = Output(np.array([1, 2]))
regex = 'The network error should be a numpy array'
with self.assertRaisesRegex(TypeError, regex):
layer.solve_bwd('void')