def test_manual_standard_logistic_regression():
X = np.array([[0, 4, 1],
[3, 0, 0],
[0, 0, 2]])
y = np.array([0, 1, 0])
Xsparse = scipy.sparse.csr_matrix(X)
for i in xrange(40):
theta = logistic.logistic_gradient_descent(X, y, max_iter=i, eta0=0.1)
thetaFast = logistic.fast_logistic_gradient_descent(X, y, max_iter=i, eta0=0.1)
thetaM, b = logistic.modified_logistic_gradient_descent(X, y, max_iter=i, b=0.0, eta0=0.1)
assert b == 0.0
thetaMFast, b = logistic.fast_modified_logistic_gradient_descent(X, y, max_iter=i, b=0.0, eta0=0.1)
assert b == 0.0
thetaSparse = logistic.fast_logistic_gradient_descent(Xsparse, y, max_iter=i, eta0=0.1)
thetaMSparse, b = logistic.fast_modified_logistic_gradient_descent(Xsparse, y, max_iter=i, b=0.0, eta0=0.1)
assert b == 0.0
close = partial(np.allclose, rtol=0.01, atol=0.0001)
'''
if i == 0:
assert np.allclose(theta, [0.01, 0.01, 0.01, 0.01])
elif i == 1:
answer = [0.0095, 0.0105, 0.0095, 0.009]
assert close(theta, answer)
theta[2] += 0.0003
assert not close(theta, answer)
theta[2] -= 0.0003
elif i == 2:
answer = [0.009, 0.011, 0.009, 0.008]
assert close(theta, answer)
assert not close(theta * 0.8, answer)
elif i == 3:
answer = [0.0085, 0.0115, 0.0085, 0.007]
assert close(theta, answer)
assert not close(theta + 0.0003, answer)
'''
assert np.allclose(theta, thetaFast)
# and make sure that the modified version is a superset of normal one
assert np.allclose(theta, thetaM)
assert np.allclose(theta, thetaMFast)
assert np.allclose(theta, thetaSparse)
assert np.allclose(theta, thetaMSparse)
for i in xrange(40):
# make sure C version is same as other version
thetaM, bM = logistic.modified_logistic_gradient_descent(X, y, max_iter=i, b=1.0, eta0=0.01)
thetaMFast, bMFast = logistic.fast_modified_logistic_gradient_descent(X, y, max_iter=i, b=1.0, eta0=0.01)
thetaMSparse, bMSparse = logistic.fast_modified_logistic_gradient_descent(Xsparse, y, max_iter=i, b=1.0, eta0=0.01)
assert np.allclose(thetaM, thetaMFast)
assert np.allclose(thetaM, thetaMSparse)
assert not np.allclose(thetaM + 0.001, thetaMFast)
assert bM == bMFast == bMSparse
def fit(self, X, y):
self.classes_, indices = unique(y, return_inverse=True)
self.theta_, self.b_ = logistic.fast_modified_logistic_gradient_descent(X, y, max_iter=self.n_iter, eta0=self.eta0, b=self.b)
return self
