def test_update_mse_should_have_the_same_effect_as_full_mse():
bounds = np.array([[50, 60, 50], [50, 60, 51], [50, 60, 52]])
target = np.random.random((100, 100, 3))
img = np.random.random((100, 100, 3))
mse = mse_full(target, img)
img[50, 50:61] = 0.5
img[51, 50:61] = 0.4
img[52, 50:61] = 0.3
update_mse(mse, bounds, img, target)
assert np.array_equal(mse, mse_full(target, img))
def test_broken_partial_mse(x1, y1, x2, y2, x3, y3, x4, y4):
img = np.zeros((500, 500, 3))
target = np.zeros((500, 500, 3))
distance = np.zeros((500, 500, 3))
assert img.dtype == np.float
assert target.dtype == np.float
assert distance.dtype == np.float
r = Quadrangle(points=np.array([[x1, y1], [x2, y2], [x3, y3], [x4, y4]],
dtype=np.int64),
alpha=0.5)
bounds = r.render(img, target)
update_mse(distance=distance, bounds=bounds, img=img, target=target)
assert np.array_equal(distance, mse_full(target, img))
assert np.mean(distance) == np.mean(mse_full(target, img))
def test_update_mse_should_calculate_the_same_values_as_full_mse_for_bounded_area(
):
mse = np.zeros((100, 100, 3))
bounds = np.array([[50, 60, 50], [50, 60, 51], [50, 60, 52]])
img = np.random.random((100, 100, 3))
target = np.random.random((100, 100, 3))
update_mse(mse, bounds, img, target)
full = mse_full(target, img)
assert np.array_equal(full[np.where(mse > 0)], mse[np.where(mse > 0)])
def get(self, X):
if self.model is None:
self.model = self.model_cls(**self.model_params)
if self.target_activations is None:
self.target_activations = self.model.get_activations(
self.canvas.target)
img_activations = self.model.get_activations(X)
target_activations = np.repeat(self.target_activations,
X.shape[0],
axis=0)
# todo: np.mean with axis=(1,2,3)
reward = self._mean(mse_full(target_activations, img_activations))
assert reward.shape[0] == X.shape[0], f'reward.shape = {reward.shape}'
return reward
def test_mse_full_should_return_float_array():
target = np.random.random((100, 100))
x = np.random.random((100, 100))
mse = mse_full(target=target, x=x)
assert mse.dtype == np.float
def test_mse_full_should_calculate_l2_properly():
target = np.array([[2, 4], [4, 3]])
x = np.array([[2, 5], [0, 2]])
mse = mse_full(target=target, x=x)
assert np.array_equal(mse, np.array([[0, 1], [16, 1]]))
def test_mse_full_should_be_max_when_target_is_white_and_x_is_black():
target = np.full((100, 100), 255)
x = np.zeros((100, 100))
mse = mse_full(target=target, x=x)
assert np.array_equal(mse, np.full(target.shape, 255 * 255))
def test_mse_full_should_be_all_zeros_when_target_and_x_are_identical():
target = np.random.random((100, 100))
x = target.copy()
mse = mse_full(target=target, x=x)
assert np.array_equal(mse, np.zeros(target.shape))