def test_grad_objective_function_infinite_end_time_two_features():
df = pd.DataFrame(
[[0, np.nan, 1, random()], [random(), np.nan, 1,
random()],
[0, random(), 1, random()], [0, random(), 1, random()]],
columns=['click_time', 'conv_time', 'intercept', 'random'])
features = ['intercept', 'random']
end_time = 1000000
y = df[['click_time', 'conv_time']].values
X = csr_matrix(df[features].values)
grad_objective_function = grad_objective_function_factory(
fit_intercept=False)
input_ = prepare_input(y, X, end_time=end_time)
input_['Jacobian'] = np.zeros(4)
y1 = input_['y1']
x0, x1 = input_['x0'], input_['x1']
w_zero = np.zeros(4)
input_['mu'] = 0
assert np.allclose(
grad_objective_function(w_zero, **input_),
np.array([
0, 1 / 2 * x0[:, 1].sum() - 1 / 2 * x1[:, 1].sum(),
(y1[:, 1] - y1[:, 0] - 1).sum(),
((y1[:, 1] - y1[:, 0] - 1) * x1[:, 1]).sum()
]))
def test_grad_objective_function_infinite_end_time():
df = pd.DataFrame([[0, np.nan, 1], [random(), np.nan, 1],
[0, random(), 1], [0, random(), 1]],
columns=['click_time', 'conv_time', 'intercept'])
end_time = 1000000
y = df[['click_time', 'conv_time']].values
X = csr_matrix(df[['intercept']].values)
grad_objective_function = grad_objective_function_factory(
fit_intercept=False)
input_ = prepare_input(y, X, end_time=end_time)
input_['Jacobian'] = np.zeros(2)
y1 = input_['y1']
w_zero = np.zeros(2)
w_one = np.ones(2)
assert np.allclose(grad_objective_function(w_zero, **input_),
np.array([0, (y1[:, 1] - y1[:, 0]).sum() - 2]))
assert np.allclose(
grad_objective_function(w_one, **input_),
np.array([
df.shape[0] * sigmoid(1) - y1.shape[0],
np.exp(1) * (y1[:, 1] - y1[:, 0]).sum() - 2
]) + w_one)
def test_grad_objective_function_output_type(test_data_input):
p = test_data_input['X'].shape[1]
grad_objective_function = grad_objective_function_factory(
fit_intercept=False)
input_ = prepare_input(**test_data_input)
w = np.zeros(2 * p)
input_['Jacobian'] = np.zeros(2 * p)
assert isinstance(grad_objective_function(w, **input_), np.ndarray)
def _get_optimization_result(self, method='L-BFGS-B', **kwargs):
p = kwargs['x0'].shape[1]
kwargs['Jacobian'] = np.zeros(2 * p)
# todo use better initialization
objective_function = objective_function_factory(self.fit_intercept)
grad_objective_function = grad_objective_function_factory(
self.fit_intercept)
def L(w):
return partial(objective_function, **kwargs)(w)
def DL(w):
return partial(grad_objective_function, **kwargs)(w)
w0 = np.zeros(2 * p)
info = minimize(L, x0=w0, method=method, jac=DL)
return info
def test_scipy_check_grad(test_dummied_matrix):
y, X = test_dummied_matrix
end_time = 1.1 * y[:, 1][~np.isnan(y[:, 1])].max()
for fit_intercept in [True, False]:
objective_function = objective_function_factory(
fit_intercept=fit_intercept)
grad_objective_function = grad_objective_function_factory(
fit_intercept=fit_intercept)
input_ = prepare_input(y, X, end_time=end_time)
p = input_['x0'].shape[1]
input_['Jacobian'] = np.zeros(2 * p)
input_['mu'] = 0.
for i in range(20):
w0 = (5. - i) * np.ones(2 * p)
L = partial(objective_function, **input_)
DL = partial(grad_objective_function, **input_)
assert isclose(check_grad(L, DL, w0), 0.)