def test_sdca_identity_poisreg(self):
"""...Test SDCA on specific case of Poisson regression with
indentity link
"""
l_l2sq = 1e-3
n_samples = 10000
n_features = 3
np.random.seed(123)
weight0 = np.random.rand(n_features)
features = np.random.rand(n_samples, n_features)
for intercept in [None, 0.45]:
if intercept is None:
fit_intercept = False
else:
fit_intercept = True
simu = SimuPoisReg(weight0, intercept=intercept, features=features,
n_samples=n_samples, link='identity',
verbose=False)
features, labels = simu.simulate()
model = ModelPoisReg(fit_intercept=fit_intercept, link='identity')
model.fit(features, labels)
sdca = SDCA(l_l2sq=l_l2sq, max_iter=100, verbose=False, tol=1e-14,
seed=Test.sto_seed)
sdca.set_model(model).set_prox(ProxZero())
start_dual = np.sqrt(sdca._rand_max * l_l2sq)
start_dual = start_dual * np.ones(sdca._rand_max)
sdca.solve(start_dual)
# Check that duality gap is 0
self.assertAlmostEqual(
sdca.objective(sdca.solution),
sdca.dual_objective(sdca.dual_solution))
# Check that original vector is approximatively retrieved
if fit_intercept:
original_coeffs = np.hstack((weight0, intercept))
else:
original_coeffs = weight0
np.testing.assert_array_almost_equal(original_coeffs,
sdca.solution, decimal=1)
# Ensure that we solve the same problem as other solvers
svrg = SVRG(max_iter=100, verbose=False, tol=1e-14,
seed=Test.sto_seed)
svrg.set_model(model).set_prox(ProxL2Sq(l_l2sq))
svrg.solve(0.5 * np.ones(model.n_coeffs), step=1e-2)
np.testing.assert_array_almost_equal(svrg.solution, sdca.solution,
decimal=4)
def test_poisreg_sdca_rand_max(self):
"""...Test that SDCA's rand_max is worrect depending on link type
"""
labels = np.array([0, 1, 2, 0, 4], dtype=self.dtype)
features = np.random.rand(len(labels), 3).astype(self.dtype)
model = ModelPoisReg(link='exponential').fit(features, labels)
with self.assertRaises(NotImplementedError):
model._sdca_rand_max
self.assertEqual(model._rand_max, 5)
model = ModelPoisReg(link='identity').fit(features, labels)
self.assertEqual(model._sdca_rand_max, 3)
self.assertEqual(model._rand_max, 5)
def prepare_solver(solver, X, y, fit_intercept=True, model="logistic",
prox="l2"):
if model == "logistic":
model = ModelLogReg(fit_intercept=fit_intercept).fit(X, y)
elif model == "poisson":
model = ModelPoisReg(fit_intercept=fit_intercept).fit(X, y)
solver.set_model(model)
if prox == "l2":
l_l2sq = TestSolver.l_l2sq
prox = ProxL2Sq(l_l2sq, (0, model.n_coeffs))
if prox is not None:
solver.set_prox(prox)
def create_model(model_type, n_samples, n_features, with_intercept=True):
weights = np.random.randn(n_features)
intercept = None
if with_intercept:
intercept = np.random.normal()
if model_type == 'Poisson':
# we need to rescale features to avoid overflows
weights /= n_features
if intercept is not None:
intercept /= n_features
if model_type == 'Linear':
simulator = SimuLinReg(weights,
intercept=intercept,
n_samples=n_samples,
verbose=False)
elif model_type == 'Logistic':
simulator = SimuLogReg(weights,
intercept=intercept,
n_samples=n_samples,
verbose=False)
elif model_type == 'Poisson':
simulator = SimuPoisReg(weights,
intercept=intercept,
n_samples=n_samples,
verbose=False)
labels, features = simulator.simulate()
if model_type == 'Linear':
model = ModelLinReg(fit_intercept=with_intercept)
elif model_type == 'Logistic':
model = ModelLogReg(fit_intercept=with_intercept)
elif model_type == 'Poisson':
model = ModelPoisReg(fit_intercept=with_intercept)
model.fit(labels, features)
return model
y = [model.loss(np.array([t])) for t in x]
plt.plot(x, y, lw=4, label=model.name)
plt.xlabel(r"$y y'$", fontsize=18)
plt.ylabel(r"$y y' \mapsto \ell(y, y')$", fontsize=18)
plt.title('Losses for binary classification', fontsize=20)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.legend(fontsize=13)
plt.subplot(1, 3, 3)
n = 1000
x = np.linspace(-1.5, 2, n)
models = [
ModelPoisReg(fit_intercept=False, link='exponential'),
ModelPoisReg(fit_intercept=False, link='identity')
]
labels = ["ModelPoisReg(link='exponential')", "ModelPoisReg(link='identity')"]
for model, label in zip(models, labels):
model.fit(np.array([[1.]]), np.array([1.]))
y = [model.loss(np.array([t])) for t in x]
plt.plot(x, y, lw=4, label=label)
plt.xlabel(r"$y'$", fontsize=16)
plt.ylabel(r"$y' \mapsto \ell(1, y')$", fontsize=16)
plt.title('Losses for count data', fontsize=20)
plt.xticks(fontsize=14)
plt.yticks(fontsize=14)
plt.legend(fontsize=13)
def test_ModelPoisReg(self):
"""...Numerical consistency check of loss and gradient for Poisson
Regression
"""
np.random.seed(12)
n_samples, n_features = 100, 10
w0 = np.random.randn(n_features) / n_features
c0 = np.random.randn() / n_features
# First check with intercept
X, y = SimuPoisReg(w0, c0, n_samples=n_samples, verbose=False,
seed=1234, dtype=self.dtype).simulate()
# Rescale features since ModelPoisReg with exponential link
# (default) is not overflow proof
X /= n_features
X_spars = csr_matrix(X, dtype=self.dtype)
model = ModelPoisReg(fit_intercept=True).fit(X, y)
model_sparse = ModelPoisReg(fit_intercept=True).fit(X_spars, y)
self.run_test_for_glm(model, model_sparse)
self._test_glm_intercept_vs_hardcoded_intercept(model)
# Then check without intercept
X, y = SimuPoisReg(w0, None, n_samples=n_samples, verbose=False,
seed=1234, dtype=self.dtype).simulate()
X /= n_features
X_spars = csr_matrix(X, dtype=self.dtype)
model = ModelPoisReg(fit_intercept=False).fit(X, y)
model_sparse = ModelPoisReg(fit_intercept=False).fit(X_spars, y)
self.run_test_for_glm(model, model_sparse)
self._test_glm_intercept_vs_hardcoded_intercept(model)
# Test the self-concordance constant
n_samples, n_features = 5, 2
X = np.zeros((n_samples, n_features))
X_spars = csr_matrix(X)
y = np.array([0, 0, 3, 2, 5], dtype=np.double)
model = ModelPoisReg(fit_intercept=True, link="identity").fit(X, y)
model_sparse = ModelPoisReg(fit_intercept=True, link="identity").fit(
X_spars, y)
self.assertAlmostEqual(model._sc_constant, 1.41421356237,
places=self.decimal_places)
self.assertAlmostEqual(model_sparse._sc_constant, 1.41421356237,
places=self.decimal_places)
y = np.array([0, 0, 3, 2, 1], dtype=np.double)
model.fit(X, y)
model_sparse.fit(X_spars, y)
self.assertAlmostEqual(model._sc_constant, 2.,
places=self.decimal_places)
self.assertAlmostEqual(model_sparse._sc_constant, 2.,
places=self.decimal_places)
def check_solver(self,
solver,
fit_intercept=True,
model='logreg',
decimal=1):
"""Check solver instance finds same parameters as scipy BFGS
Parameters
----------
solver : `Solver`
Instance of a solver to be tested
fit_intercept : `bool`, default=True
Model uses intercept is `True`
model : 'linreg' | 'logreg' | 'poisreg', default='logreg'
Name of the model used to test the solver
decimal : `int`, default=1
Number of decimals required for the test
"""
# Set seed for data simulation
np.random.seed(12)
n_samples = TestSolver.n_samples
n_features = TestSolver.n_features
coeffs0 = weights_sparse_gauss(n_features, nnz=5)
if fit_intercept:
interc0 = 2.
else:
interc0 = None
if model == 'linreg':
X, y = SimuLinReg(coeffs0,
interc0,
n_samples=n_samples,
verbose=False,
seed=123).simulate()
model = ModelLinReg(fit_intercept=fit_intercept).fit(X, y)
elif model == 'logreg':
X, y = SimuLogReg(coeffs0,
interc0,
n_samples=n_samples,
verbose=False,
seed=123).simulate()
model = ModelLogReg(fit_intercept=fit_intercept).fit(X, y)
elif model == 'poisreg':
X, y = SimuPoisReg(coeffs0,
interc0,
n_samples=n_samples,
verbose=False,
seed=123).simulate()
# Rescale features to avoid overflows in Poisson simulations
X /= np.linalg.norm(X, axis=1).reshape(n_samples, 1)
model = ModelPoisReg(fit_intercept=fit_intercept).fit(X, y)
else:
raise ValueError("``model`` must be either 'linreg', 'logreg' or"
" 'poisreg'")
solver.set_model(model)
strength = 1e-2
prox = ProxL2Sq(strength, (0, model.n_features))
if type(solver) is not SDCA:
solver.set_prox(prox)
else:
solver.set_prox(ProxZero())
solver.l_l2sq = strength
coeffs_solver = solver.solve()
# Compare with BFGS
bfgs = BFGS(max_iter=100,
verbose=False).set_model(model).set_prox(prox)
coeffs_bfgs = bfgs.solve()
np.testing.assert_almost_equal(coeffs_solver,
coeffs_bfgs,
decimal=decimal)
# We ensure that reached coeffs are not equal to zero
self.assertGreater(norm(coeffs_solver), 0)
self.assertAlmostEqual(solver.objective(coeffs_bfgs),
solver.objective(coeffs_solver),
delta=1e-2)
