def test_grad_search(model, crit):
"""check that the paths are the same in the line search"""
n_outer = 2
criterion = HeldOutLogistic(X_val, y_val, model)
monitor1 = Monitor()
algo = Forward()
grad_search(algo, criterion, log_alpha, monitor1, n_outer=n_outer,
tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model)
monitor2 = Monitor()
algo = Implicit()
grad_search(algo, criterion, log_alpha, monitor2, n_outer=n_outer,
tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model)
monitor3 = Monitor()
algo = ImplicitForward(tol_jac=tol, n_iter_jac=5000)
grad_search(algo, criterion, log_alpha, monitor3, n_outer=n_outer,
tol=tol)
assert np.allclose(
np.array(monitor1.log_alphas), np.array(monitor3.log_alphas))
assert np.allclose(
np.array(monitor1.grads), np.array(monitor3.grads), atol=1e-4)
assert np.allclose(
np.array(monitor1.objs), np.array(monitor3.objs))
assert not np.allclose(
np.array(monitor1.times), np.array(monitor3.times))
def test_val_grad(model):
#######################################################################
# Not all methods computes the full Jacobian, but all
# compute the gradients
# check that the gradient returned by all methods are the same
criterion = HeldOutLogistic(X_val, y_val, model)
algo = Forward()
val_fwd, grad_fwd = criterion.get_val_grad(log_C,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=100)
val_imp_fwd, grad_imp_fwd = criterion.get_val_grad(log_C,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model)
algo = Implicit()
val_imp, grad_imp = criterion.get_val_grad(log_C,
algo.get_beta_jac_v,
tol=tol)
assert np.allclose(val_fwd, val_imp_fwd)
assert np.allclose(grad_fwd, grad_imp_fwd)
assert np.allclose(val_imp_fwd, val_imp)
assert np.allclose(grad_imp_fwd, grad_imp, atol=1e-5)
def test_val_grad(model):
criterion = HeldOutLogistic(idx_val, idx_val)
algo = Forward()
val_fwd, grad_fwd = criterion.get_val_grad(model,
X,
y,
log_alpha,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(idx_val, idx_val)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000)
val_imp_fwd, grad_imp_fwd = criterion.get_val_grad(model,
X,
y,
log_alpha,
algo.get_beta_jac_v,
tol=tol)
criterion = HeldOutLogistic(idx_val, idx_val)
algo = Implicit()
val_imp, grad_imp = criterion.get_val_grad(model,
X,
y,
log_alpha,
algo.get_beta_jac_v,
tol=tol)
assert np.allclose(val_fwd, val_imp_fwd, atol=1e-4)
assert np.allclose(grad_fwd, grad_imp_fwd, atol=1e-4)
assert np.allclose(val_imp_fwd, val_imp, atol=1e-4)
# for the implcit the conjugate grad does not converge
# hence the rtol=1e-2
assert np.allclose(grad_imp_fwd, grad_imp, rtol=1e-2)
def test_check_grad_sparse_ho(model_name, criterion, algo):
"""Check that all methods return a good gradient using check_grad"""
if criterion == 'MSE':
criterion = HeldOutMSE(idx_train, idx_val)
elif criterion == 'SURE':
criterion = FiniteDiffMonteCarloSure(sigma_star)
elif criterion == 'logistic':
criterion = HeldOutLogistic(idx_train, idx_val)
model = models[model_name]
log_alpha = dict_log_alpha[model_name]
def get_val(log_alpha):
val, _ = criterion.get_val_grad(
model, X, y, np.squeeze(log_alpha), algo.get_beta_jac_v, tol=tol)
return val
def get_grad(log_alpha):
_, grad = criterion.get_val_grad(
model, X, y, np.squeeze(log_alpha), algo.get_beta_jac_v, tol=tol)
return grad
for log_alpha in dict_list_log_alphas[model_name]:
grad_error = check_grad(get_val, get_grad, log_alpha)
assert grad_error < 1e-1
def test_val_grad(model_name, criterion_name, algo):
"""Check that all methods return the same gradient, comparing to cvxpylayer
"""
if model_name == 'svr':
pytest.xfail("svr needs to be fixed")
if criterion_name == 'logistic':
pytest.xfail("cvxpylayer seems broken for logistic")
if criterion_name == 'MSE':
criterion = HeldOutMSE(idx_train, idx_val)
elif criterion_name == 'logistic':
criterion = HeldOutLogistic(idx_train, idx_val)
elif criterion_name == 'SURE':
criterion = FiniteDiffMonteCarloSure(sigma_star)
log_alpha = dict_log_alpha[model_name]
model = models[model_name]
val, grad = criterion.get_val_grad(model,
X,
y,
log_alpha,
algo.compute_beta_grad,
tol=tol)
np.testing.assert_allclose(dict_vals_cvxpy[model_name, criterion_name],
val,
rtol=1e-5,
atol=1e-5)
np.testing.assert_allclose(dict_grads_cvxpy[model_name, criterion_name],
grad,
rtol=1e-5,
atol=1e-5)
def test_grad_search_custom(model, model_custom, crit):
"""check that the paths are the same in the line search"""
n_outer = 5
criterion = HeldOutLogistic(idx_val, idx_val)
monitor = Monitor()
algo = ImplicitForward(tol_jac=tol, n_iter_jac=5000)
grad_search(algo,
criterion,
model,
X,
y,
log_alpha,
monitor,
n_outer=n_outer,
tol=tol)
criterion = HeldOutLogistic(idx_val, idx_val)
monitor_custom = Monitor()
algo = ImplicitForward(tol_jac=tol, n_iter_jac=5000)
grad_search(algo,
criterion,
model_custom,
X,
y,
log_alpha,
monitor_custom,
n_outer=n_outer,
tol=tol)
np.testing.assert_allclose(np.array(monitor.log_alphas),
np.array(monitor_custom.log_alphas),
atol=1e-3)
np.testing.assert_allclose(np.array(monitor.grads),
np.array(monitor_custom.grads),
atol=1e-4)
np.testing.assert_allclose(np.array(monitor.objs),
np.array(monitor_custom.objs),
atol=1e-5)
assert not np.allclose(np.array(monitor.times),
np.array(monitor_custom.times))
def test_cross_val_criterion(model_name, XX):
model = models[model_name]
alpha_min = alpha_max / 10
max_iter = 10000
n_alphas = 10
kf = KFold(n_splits=5, shuffle=True, random_state=56)
monitor_grid = Monitor()
if model_name.startswith("lasso"):
sub_crit = HeldOutMSE(None, None)
else:
sub_crit = HeldOutLogistic(None, None)
criterion = CrossVal(sub_crit, cv=kf)
grid_search(criterion,
model,
XX,
y,
alpha_min,
alpha_max,
monitor_grid,
max_evals=n_alphas,
tol=tol)
if model_name.startswith("lasso"):
reg = linear_model.LassoCV(cv=kf,
verbose=True,
tol=tol,
fit_intercept=False,
alphas=np.geomspace(alpha_max,
alpha_min,
num=n_alphas),
max_iter=max_iter).fit(X, y)
else:
reg = linear_model.LogisticRegressionCV(
cv=kf,
verbose=True,
tol=tol,
fit_intercept=False,
Cs=len(idx_train) /
np.geomspace(alpha_max, alpha_min, num=n_alphas),
max_iter=max_iter,
penalty='l1',
solver='liblinear').fit(X, y)
reg.score(XX, y)
if model_name.startswith("lasso"):
objs_grid_sk = reg.mse_path_.mean(axis=1)
else:
objs_grid_sk = reg.scores_[1.0].mean(axis=1)
# these 2 value should be the same
(objs_grid_sk - np.array(monitor_grid.objs))
np.testing.assert_allclose(objs_grid_sk, monitor_grid.objs)
def test_val_grad_custom(model, model_custom):
criterion = HeldOutLogistic(X_val, y_val, model)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000)
val, grad = criterion.get_val_grad(
log_alpha, algo.get_beta_jac_v, tol=tol)
criterion = HeldOutLogistic(X_val, y_val, model_custom)
algo = ImplicitForward(tol_jac=1e-8, n_iter_jac=5000)
val_custom, grad_custom = criterion.get_val_grad(
log_alpha, algo.get_beta_jac_v, tol=tol)
assert np.allclose(val, val_custom)
assert np.allclose(grad, grad_custom)
p_alphas = np.geomspace(1, 0.0001, n_alphas)
alphas = alpha_max * p_alphas
log_alphas = np.log(alphas)
##############################################################################
# Grid-search
# -----------
print('scikit started')
t0 = time.time()
estimator = LogisticRegression(penalty='l1',
fit_intercept=False,
max_iter=max_iter)
model = SparseLogreg(max_iter=max_iter, estimator=estimator)
criterion = HeldOutLogistic(idx_train, idx_val)
algo_grid = Forward()
monitor_grid = Monitor()
grid_search(algo_grid,
criterion,
model,
X,
y,
log_alpha_min,
log_alpha_max,
monitor_grid,
log_alphas=log_alphas,
tol=tol)
objs = np.array(monitor_grid.objs)
t_sk = time.time() - t0
alphas = alpha_max * p_alphas
log_alphas = np.log(alphas)
##############################################################################
# Grid-search
# -----------
print('scikit started')
t0 = time.time()
estimator = LogisticRegression(penalty='l1',
fit_intercept=False,
solver='saga',
max_iter=max_iter)
model = SparseLogreg(X_train, y_train, max_iter=max_iter, estimator=estimator)
criterion = HeldOutLogistic(X_val, y_val, model)
algo_grid = Forward()
monitor_grid = Monitor()
grid_search(algo_grid,
criterion,
log_alpha_min,
log_alpha_max,
monitor_grid,
log_alphas=log_alphas,
tol=tol)
objs = np.array(monitor_grid.objs)
t_sk = time.time() - t0
print('scikit finished')
print("Time to compute CV for scikit-learn: %.2f" % t_sk)
def parallel_function(dataset_name,
method,
tol=1e-5,
n_outer=50,
tolerance_decrease='constant'):
# load data
X, y = fetch_libsvm(dataset_name)
y -= np.mean(y)
# compute alpha_max
alpha_max = np.abs(X.T @ y).max() / len(y)
if model_name == "logreg":
alpha_max /= 2
alpha_min = alpha_max * dict_palphamin[dataset_name]
if model_name == "enet":
estimator = linear_model.ElasticNet(fit_intercept=False,
max_iter=10_000,
warm_start=True,
tol=tol)
model = ElasticNet(estimator=estimator)
elif model_name == "logreg":
model = SparseLogreg(estimator=estimator)
# TODO improve this
try:
n_outer = dict_n_outers[dataset_name, method]
except Exception:
n_outer = 20
size_loop = 2
for _ in range(size_loop):
if model_name == "lasso" or model_name == "enet":
sub_criterion = HeldOutMSE(None, None)
elif model_name == "logreg":
criterion = HeldOutLogistic(None, None)
kf = KFold(n_splits=5, shuffle=True, random_state=42)
criterion = CrossVal(sub_criterion, cv=kf)
algo = ImplicitForward(tol_jac=1e-3)
monitor = Monitor()
t_max = dict_t_max[dataset_name]
if method == 'grid_search':
num1D = dict_point_grid_search[dataset_name]
alpha1D = np.geomspace(alpha_max, alpha_min, num=num1D)
alphas = [np.array(i) for i in product(alpha1D, alpha1D)]
grid_search(algo,
criterion,
model,
X,
y,
alpha_min,
alpha_max,
monitor,
max_evals=100,
tol=tol,
alphas=alphas)
elif method == 'random' or method == 'bayesian':
hyperopt_wrapper(algo,
criterion,
model,
X,
y,
alpha_min,
alpha_max,
monitor,
max_evals=30,
tol=tol,
method=method,
size_space=2,
t_max=t_max)
elif method.startswith("implicit_forward"):
# do gradient descent to find the optimal lambda
alpha0 = np.array([alpha_max / 100, alpha_max / 100])
n_outer = 30
if method == 'implicit_forward':
optimizer = GradientDescent(n_outer=n_outer,
p_grad_norm=1,
verbose=True,
tol=tol,
t_max=t_max)
else:
optimizer = GradientDescent(n_outer=n_outer,
p_grad_norm=1,
verbose=True,
tol=tol,
t_max=t_max,
tol_decrease="geom")
grad_search(algo, criterion, model, optimizer, X, y, alpha0,
monitor)
else:
raise NotImplementedError
monitor.times = np.array(monitor.times)
monitor.objs = np.array(monitor.objs)
monitor.objs_test = 0 # TODO
monitor.alphas = np.array(monitor.alphas)
return (dataset_name, method, tol, n_outer, tolerance_decrease,
monitor.times, monitor.objs, monitor.objs_test, monitor.alphas,
alpha_max, model_name)
import pytest
import numpy as np
from sparse_ho.criterion import (HeldOutMSE, HeldOutLogistic,
FiniteDiffMonteCarloSure)
from sparse_ho.utils import Monitor
from sparse_ho import Forward
from sparse_ho.tests.common import (X, y, sigma_star, idx_train, idx_val,
models, dict_list_log_alphas)
list_model_crit = [('lasso', HeldOutMSE(idx_train, idx_val)),
('enet', HeldOutMSE(idx_train, idx_val)),
('wLasso', HeldOutMSE(idx_train, idx_val)),
('lasso', FiniteDiffMonteCarloSure(sigma_star)),
('logreg', HeldOutLogistic(idx_train, idx_val))]
tol = 1e-15
@pytest.mark.parametrize('model_name,criterion', list_model_crit)
def test_cross_val_criterion(model_name, criterion):
# verify dtype from criterion, and the good shape
algo = Forward()
monitor_get_val = Monitor()
monitor_get_val_grad = Monitor()
model = models[model_name]
for log_alpha in dict_list_log_alphas[model_name]:
criterion.get_val(model,
X,
# X_train, X_val, X_test, y_train, y_val, y_test = get_leukemia()
n_samples, n_features = X_train.shape
print("Starting path computation...")
alpha_max = np.max(np.abs(X_train.T @ (-y_train)))
alpha_max /= (2 * n_samples)
n_alphas = 10
p_alphas = np.geomspace(1, 1e-4, n_alphas)
alphas = p_alphas * alpha_max
log_alphas = np.log(alphas)
tol = 1e-5
# grid search
model = SparseLogreg(X_train, y_train, log_alphas[0], max_iter=1000)
criterion = HeldOutLogistic(X_val, y_val, model, X_test=X_test, y_test=y_test)
algo = Forward(criterion)
monitor_grid_sk = Monitor()
grid_search(algo, None, None, monitor_grid_sk, log_alphas=log_alphas, tol=tol)
monitor = Monitor()
# grad_search(
# algo, logCs[0], monitor, n_outer=5, verbose=True,
# tolerance_decrease='constant', tol=1e-8,
# t_max=10000)
plt.figure()
plt.plot(monitor_grid_sk.log_alphas, monitor_grid_sk.objs)
plt.plot(monitor.log_alphas, monitor.objs, 'bo')
plt.show(block=False)
def parallel_function(dataset_name,
method,
tol=1e-5,
n_outer=50,
tolerance_decrease='constant'):
# load data
X_train, X_val, X_test, y_train, y_val, y_test = get_data(dataset_name)
n_samples, _ = X_train.shape
# compute alpha_max
alpha_max = np.abs(X_train.T @ y_train).max() / n_samples
if model_name == "logreg":
alpha_max /= 2
alpha_min = alpha_max / 10_000
log_alpha_max = np.log(alpha_max)
log_alpha_min = np.log(alpha_min)
log_alpha0 = np.log(0.1 * alpha_max)
if model_name == "lasso":
model = Lasso(X_train, y_train)
elif model_name == "logreg":
model = SparseLogreg(X_train, y_train)
try:
n_outer = dict_n_outers[dataset_name, method]
except Exception:
n_outer = 20
size_loop = 2
for _ in range(size_loop):
if model_name == "lasso":
criterion = HeldOutMSE(X_val,
y_val,
model,
X_test=X_test,
y_test=y_test)
elif model_name == "logreg":
criterion = HeldOutLogistic(X_val,
y_val,
model,
X_test=X_test,
y_test=y_test)
algo = dict_algo[method](criterion)
monitor = Monitor()
if method == 'grid_search':
log_alphas = np.log(np.geomspace(alpha_max, alpha_min, num=100))
grid_search(algo,
None,
None,
monitor,
log_alphas=log_alphas,
tol=tol)
elif method == 'random':
grid_search(algo,
log_alpha_max,
log_alpha_min,
monitor,
tol=tol,
max_evals=n_alphas,
t_max=dict_t_max[dataset_name])
elif method in ("bayesian"):
hyperopt_wrapper(algo,
log_alpha_min,
log_alpha_max,
monitor,
max_evals=n_alphas,
tol=tol,
method='bayesian',
t_max=dict_t_max[dataset_name])
else:
# do line search to find the optimal lambda
grad_search(algo,
log_alpha0,
monitor,
n_outer=n_outer,
tol=tol,
tolerance_decrease=tolerance_decrease,
t_max=dict_t_max[dataset_name])
monitor.times = np.array(monitor.times)
monitor.objs = np.array(monitor.objs)
monitor.objs_test = np.array(monitor.objs_test)
monitor.log_alphas = np.array(monitor.log_alphas)
return (dataset_name, method, tol, n_outer, tolerance_decrease,
monitor.times, monitor.objs, monitor.objs_test, monitor.log_alphas,
norm(y_val), norm(y_test), log_alpha_max, model_name)
def parallel_function(
dataset_name, method, tol=1e-5, n_outer=50,
tolerance_decrease='exponential'):
# load data
X_train, X_val, X_test, y_train, y_val, y_test = get_data(dataset_name)
n_samples, n_features = X_train.shape
print('n_samples', n_samples)
print('n_features', n_features)
y_train[y_train == 0.0] = -1.0
y_val[y_val == 0.0] = -1.0
y_test[y_test == 0.0] = -1.0
alpha_max = np.max(np.abs(X_train.T @ y_train))
alpha_max /= X_train.shape[0]
alpha_max /= 4
log_alpha_max = np.log(alpha_max)
alpha_min = alpha_max * 1e-4
alphas = np.geomspace(alpha_max, alpha_min, 10)
log_alphas = np.log(alphas)
log_alpha0 = np.log(0.1 * alpha_max)
log_alpha_max = np.log(alpha_max)
n_outer = 25
if dataset_name == "rcv1":
size_loop = 2
else:
size_loop = 2
model = SparseLogreg(
X_train, y_train, max_iter=1000, log_alpha_max=log_alpha_max)
for i in range(size_loop):
monitor = Monitor()
if method == "implicit_forward":
criterion = HeldOutLogistic(X_val, y_val, model, X_test=X_test, y_test=y_test)
algo = ImplicitForward(criterion, tol_jac=1e-5, n_iter_jac=100)
_, _, _ = grad_search(
algo=algo, verbose=False,
log_alpha0=log_alpha0, tol=tol,
n_outer=n_outer, monitor=monitor,
t_max=dict_t_max[dataset_name],
tolerance_decrease=tolerance_decrease)
elif method == "forward":
criterion = HeldOutLogistic(X_val, y_val, model, X_test=X_test, y_test=y_test)
algo = Forward(criterion)
_, _, _ = grad_search(
algo=algo,
log_alpha0=log_alpha0, tol=tol,
n_outer=n_outer, monitor=monitor,
t_max=dict_t_max[dataset_name],
tolerance_decrease=tolerance_decrease)
elif method == "implicit":
criterion = HeldOutLogistic(X_val, y_val, model, X_test=X_test, y_test=y_test)
algo = Implicit(criterion)
_, _, _ = grad_search(
algo=algo,
log_alpha0=log_alpha0, tol=tol,
n_outer=n_outer, monitor=monitor,
t_max=dict_t_max[dataset_name],
tolerance_decrease=tolerance_decrease)
elif method == "grid_search":
criterion = HeldOutLogistic(X_val, y_val, model, X_test=X_test, y_test=y_test)
algo = Forward(criterion)
# log_alpha_min = np.log(alpha_min)
log_alphas = np.log(np.geomspace(alpha_max, alpha_min, num=100))
log_alpha_opt, min_g_func = grid_search(
algo, None, None, monitor, tol=tol, samp="grid",
t_max=dict_t_max[dataset_name], log_alphas=log_alphas)
print(log_alpha_opt)
elif method == "random":
criterion = HeldOutLogistic(X_val, y_val, model, X_test=X_test, y_test=y_test)
algo = Forward(criterion)
log_alpha_min = np.log(alpha_min)
log_alpha_opt, min_g_func = grid_search(
algo, log_alpha_min, np.log(alpha_max), monitor, max_evals=100, tol=tol, samp="random",
t_max=dict_t_max[dataset_name])
print(log_alpha_opt)
elif method == "lhs":
criterion = HeldOutLogistic(X_val, y_val, model, X_test=X_test, y_test=y_test)
algo = Forward(criterion)
log_alpha_min = np.log(alpha_min)
log_alpha_opt, min_g_func = grid_search(
algo, log_alpha_min, np.log(alpha_max), monitor, max_evals=100, tol=tol, samp="lhs",
t_max=dict_t_max[dataset_name])
print(log_alpha_opt)
monitor.times = np.array(monitor.times).copy()
monitor.objs = np.array(monitor.objs).copy()
monitor.objs_test = np.array(monitor.objs_test).copy()
monitor.log_alphas = np.array(monitor.log_alphas).copy()
return (dataset_name, method, tol, n_outer, tolerance_decrease,
monitor.times, monitor.objs, monitor.objs_test,
monitor.log_alphas, norm(y_val), norm(y_test), log_alpha_max)
