def _logistic_regression_path(X,
y,
pos_class=None,
Cs=10,
fit_intercept=True,
max_iter=100,
tol=1e-4,
verbose=0,
solver='lbfgs',
coef=None,
class_weight=None,
dual=False,
penalty='l2',
intercept_scaling=1.,
multi_class='warn',
random_state=None,
check_input=True,
max_squared_sum=None,
sample_weight=None,
l1_ratio=None):
"""Compute a Logistic Regression model for a list of regularization
parameters.
This is an implementation that uses the result of the previous model
to speed up computations along the set of solutions, making it faster
than sequentially calling LogisticRegression for the different parameters.
Note that there will be no speedup with liblinear solver, since it does
not handle warm-starting.
Read more in the :ref:`User Guide <logistic_regression>`.
Parameters
----------
X : array-like or sparse matrix, shape (n_samples, n_features)
Input data.
y : array-like, shape (n_samples,) or (n_samples, n_targets)
Input data, target values.
pos_class : int, None
The class with respect to which we perform a one-vs-all fit.
If None, then it is assumed that the given problem is binary.
Cs : int | array-like, shape (n_cs,)
List of values for the regularization parameter or integer specifying
the number of regularization parameters that should be used. In this
case, the parameters will be chosen in a logarithmic scale between
1e-4 and 1e4.
fit_intercept : bool
Whether to fit an intercept for the model. In this case the shape of
the returned array is (n_cs, n_features + 1).
max_iter : int
Maximum number of iterations for the solver.
tol : float
Stopping criterion. For the newton-cg and lbfgs solvers, the iteration
will stop when ``max{|g_i | i = 1, ..., n} <= tol``
where ``g_i`` is the i-th component of the gradient.
verbose : int
For the liblinear and lbfgs solvers set verbose to any positive
number for verbosity.
solver : {'lbfgs', 'newton-cg', 'liblinear', 'sag', 'saga'}
Numerical solver to use.
coef : array-like, shape (n_features,), default None
Initialization value for coefficients of logistic regression.
Useless for liblinear solver.
class_weight : dict or 'balanced', optional
Weights associated with classes in the form ``{class_label: weight}``.
If not given, all classes are supposed to have weight one.
The "balanced" mode uses the values of y to automatically adjust
weights inversely proportional to class frequencies in the input data
as ``n_samples / (n_classes * np.bincount(y))``.
Note that these weights will be multiplied with sample_weight (passed
through the fit method) if sample_weight is specified.
dual : bool
Dual or primal formulation. Dual formulation is only implemented for
l2 penalty with liblinear solver. Prefer dual=False when
n_samples > n_features.
penalty : str, 'l1', 'l2', or 'elasticnet'
Used to specify the norm used in the penalization. The 'newton-cg',
'sag' and 'lbfgs' solvers support only l2 penalties. 'elasticnet' is
only supported by the 'saga' solver.
intercept_scaling : float, default 1.
Useful only when the solver 'liblinear' is used
and self.fit_intercept is set to True. In this case, x becomes
[x, self.intercept_scaling],
i.e. a "synthetic" feature with constant value equal to
intercept_scaling is appended to the instance vector.
The intercept becomes ``intercept_scaling * synthetic_feature_weight``.
Note! the synthetic feature weight is subject to l1/l2 regularization
as all other features.
To lessen the effect of regularization on synthetic feature weight
(and therefore on the intercept) intercept_scaling has to be increased.
multi_class : str, {'ovr', 'multinomial', 'auto'}, default: 'ovr'
If the option chosen is 'ovr', then a binary problem is fit for each
label. For 'multinomial' the loss minimised is the multinomial loss fit
across the entire probability distribution, *even when the data is
binary*. 'multinomial' is unavailable when solver='liblinear'.
'auto' selects 'ovr' if the data is binary, or if solver='liblinear',
and otherwise selects 'multinomial'.
.. versionadded:: 0.18
Stochastic Average Gradient descent solver for 'multinomial' case.
.. versionchanged:: 0.20
Default will change from 'ovr' to 'auto' in 0.22.
random_state : int, RandomState instance or None, optional, default None
The seed of the pseudo random number generator to use when shuffling
the data. If int, random_state is the seed used by the random number
generator; If RandomState instance, random_state is the random number
generator; If None, the random number generator is the RandomState
instance used by `np.random`. Used when ``solver`` == 'sag' or
'liblinear'.
check_input : bool, default True
If False, the input arrays X and y will not be checked.
max_squared_sum : float, default None
Maximum squared sum of X over samples. Used only in SAG solver.
If None, it will be computed, going through all the samples.
The value should be precomputed to speed up cross validation.
sample_weight : array-like, shape(n_samples,) optional
Array of weights that are assigned to individual samples.
If not provided, then each sample is given unit weight.
l1_ratio : float or None, optional (default=None)
The Elastic-Net mixing parameter, with ``0 <= l1_ratio <= 1``. Only
used if ``penalty='elasticnet'``. Setting ``l1_ratio=0`` is equivalent
to using ``penalty='l2'``, while setting ``l1_ratio=1`` is equivalent
to using ``penalty='l1'``. For ``0 < l1_ratio <1``, the penalty is a
combination of L1 and L2.
Returns
-------
coefs : ndarray, shape (n_cs, n_features) or (n_cs, n_features + 1)
List of coefficients for the Logistic Regression model. If
fit_intercept is set to True then the second dimension will be
n_features + 1, where the last item represents the intercept. For
``multiclass='multinomial'``, the shape is (n_classes, n_cs,
n_features) or (n_classes, n_cs, n_features + 1).
Cs : ndarray
Grid of Cs used for cross-validation.
n_iter : array, shape (n_cs,)
Actual number of iteration for each Cs.
Notes
-----
You might get slightly different results with the solver liblinear than
with the others since this uses LIBLINEAR which penalizes the intercept.
.. versionchanged:: 0.19
The "copy" parameter was removed.
"""
if isinstance(Cs, numbers.Integral):
Cs = np.logspace(-4, 4, Cs)
solver = _check_solver(solver, penalty, dual)
# Preprocessing.
if check_input:
X = check_array(X,
accept_sparse='csr',
dtype=np.float64,
accept_large_sparse=solver != 'liblinear')
y = check_array(y, ensure_2d=False, dtype=None)
check_consistent_length(X, y)
_, n_features = X.shape
classes = np.unique(y)
random_state = check_random_state(random_state)
multi_class = _check_multi_class(multi_class, solver, len(classes))
if pos_class is None and multi_class != 'multinomial':
if (classes.size > 2):
raise ValueError('To fit OvR, use the pos_class argument')
# np.unique(y) gives labels in sorted order.
pos_class = classes[1]
# If sample weights exist, convert them to array (support for lists)
# and check length
# Otherwise set them to 1 for all examples
if sample_weight is not None:
sample_weight = np.array(sample_weight, dtype=X.dtype, order='C')
check_consistent_length(y, sample_weight)
default_weights = False
else:
sample_weight = np.ones(X.shape[0], dtype=X.dtype)
default_weights = (class_weight is None)
daal_ready = use_daal and solver in ['lbfgs', 'newton-cg'
] and not sparse.issparse(X)
# If class_weights is a dict (provided by the user), the weights
# are assigned to the original labels. If it is "balanced", then
# the class_weights are assigned after masking the labels with a OvR.
le = LabelEncoder()
if isinstance(class_weight, dict) or multi_class == 'multinomial':
class_weight_ = compute_class_weight(class_weight, classes, y)
sample_weight *= class_weight_[le.fit_transform(y)]
# For doing a ovr, we need to mask the labels first. for the
# multinomial case this is not necessary.
if multi_class == 'ovr':
w0 = np.zeros(n_features + int(fit_intercept), dtype=X.dtype)
mask_classes = np.array([-1, 1])
mask = (y == pos_class)
y_bin = np.ones(y.shape, dtype=X.dtype)
y_bin[~mask] = -1.
# for compute_class_weight
if class_weight == "balanced":
class_weight_ = compute_class_weight(class_weight, mask_classes,
y_bin)
sample_weight *= class_weight_[le.fit_transform(y_bin)]
daal_ready = daal_ready and (default_weights or np.allclose(
sample_weight, np.ones_like(sample_weight)))
if daal_ready:
w0 = np.zeros(n_features + 1, dtype=X.dtype)
y_bin[~mask] = 0.
else:
w0 = np.zeros(n_features + int(fit_intercept), dtype=X.dtype)
else:
daal_ready = daal_ready and (default_weights or np.allclose(
sample_weight, np.ones_like(sample_weight)))
if solver not in ['sag', 'saga']:
if daal_ready:
Y_multi = le.fit_transform(y).astype(X.dtype, copy=False)
else:
lbin = LabelBinarizer()
Y_multi = lbin.fit_transform(y)
if Y_multi.shape[1] == 1:
Y_multi = np.hstack([1 - Y_multi, Y_multi])
else:
# SAG multinomial solver needs LabelEncoder, not LabelBinarizer
le = LabelEncoder()
Y_multi = le.fit_transform(y).astype(X.dtype, copy=False)
if daal_ready:
w0 = np.zeros((classes.size, n_features + 1),
order='C',
dtype=X.dtype)
else:
w0 = np.zeros((classes.size, n_features + int(fit_intercept)),
order='F',
dtype=X.dtype)
if coef is not None:
# it must work both giving the bias term and not
if multi_class == 'ovr':
if coef.size not in (n_features, w0.size):
raise ValueError(
'Initialization coef is of shape %d, expected shape '
'%d or %d' % (coef.size, n_features, w0.size))
if daal_ready:
w0[-coef.size:] = np.roll(
coef, 1, -1) if coef.size != n_features else coef
else:
w0[:coef.size] = coef
else:
# For binary problems coef.shape[0] should be 1, otherwise it
# should be classes.size.
n_classes = classes.size
if n_classes == 2:
n_classes = 1
if (coef.shape[0] != n_classes
or coef.shape[1] not in (n_features, n_features + 1)):
raise ValueError(
'Initialization coef is of shape (%d, %d), expected '
'shape (%d, %d) or (%d, %d)' %
(coef.shape[0], coef.shape[1], classes.size, n_features,
classes.size, n_features + 1))
if daal_ready:
w0[:, -coef.shape[1]:] = np.roll(
coef, 1, -1) if coef.shape[1] != n_features else coef
else:
if n_classes == 1:
w0[0, :coef.shape[1]] = -coef
w0[1, :coef.shape[1]] = coef
else:
w0[:, :coef.shape[1]] = coef
C_daal_multiplier = 1
# commented out because this is Py3 feature
#def _map_to_binary_logistic_regression():
# nonlocal C_daal_multiplier
# nonlocal w0
# C_daal_multiplier = 2
# w0 *= 2
if multi_class == 'multinomial':
# fmin_l_bfgs_b and newton-cg accepts only ravelled parameters.
if solver in ['lbfgs', 'newton-cg']:
if daal_ready and classes.size == 2:
w0_saved = w0
w0 = w0[-1:, :]
w0 = w0.ravel()
target = Y_multi
if solver == 'lbfgs':
if daal_ready:
if classes.size == 2:
# _map_to_binary_logistic_regression()
C_daal_multiplier = 2
w0 *= 2
daal_extra_args_func = _daal4py_logistic_loss_extra_args
else:
daal_extra_args_func = _daal4py_cross_entropy_loss_extra_args
func = _daal4py_loss_and_grad
else:
func = lambda x, *args: _multinomial_loss_grad(x, *args)[0:2]
elif solver == 'newton-cg':
if daal_ready:
if classes.size == 2:
# _map_to_binary_logistic_regression()
C_daal_multiplier = 2
w0 *= 2
daal_extra_args_func = _daal4py_logistic_loss_extra_args
else:
daal_extra_args_func = _daal4py_cross_entropy_loss_extra_args
func = _daal4py_loss_
grad = _daal4py_grad_
hess = _daal4py_grad_hess_
else:
func = lambda x, *args: _multinomial_loss(x, *args)[0]
grad = lambda x, *args: _multinomial_loss_grad(x, *args)[1]
hess = _multinomial_grad_hess
warm_start_sag = {'coef': w0.T}
else:
target = y_bin
if solver == 'lbfgs':
if daal_ready:
func = _daal4py_loss_and_grad
daal_extra_args_func = _daal4py_logistic_loss_extra_args
else:
func = _logistic_loss_and_grad
elif solver == 'newton-cg':
if daal_ready:
daal_extra_args_func = _daal4py_logistic_loss_extra_args
func = _daal4py_loss_
grad = _daal4py_grad_
hess = _daal4py_grad_hess_
else:
func = _logistic_loss
grad = lambda x, *args: _logistic_loss_and_grad(x, *args)[1]
hess = _logistic_grad_hess
warm_start_sag = {'coef': np.expand_dims(w0, axis=1)}
coefs = list()
n_iter = np.zeros(len(Cs), dtype=np.int32)
for i, C in enumerate(Cs):
if solver == 'lbfgs':
if daal_ready:
extra_args = daal_extra_args_func(classes.size,
w0,
X,
target,
0.,
0.5 / C / C_daal_multiplier,
fit_intercept,
value=True,
gradient=True,
hessian=False)
else:
extra_args = (X, target, 1. / C, sample_weight)
iprint = [-1, 50, 1, 100,
101][np.searchsorted(np.array([0, 1, 2, 3]), verbose)]
w0, loss, info = optimize.fmin_l_bfgs_b(func,
w0,
fprime=None,
args=extra_args,
iprint=iprint,
pgtol=tol,
maxiter=max_iter)
if daal_ready and C_daal_multiplier == 2:
w0 *= 0.5
if info["warnflag"] == 1:
warnings.warn(
"lbfgs failed to converge. Increase the number "
"of iterations.", ConvergenceWarning)
# In scipy <= 1.0.0, nit may exceed maxiter.
# See https://github.com/scipy/scipy/issues/7854.
n_iter_i = min(info['nit'], max_iter)
elif solver == 'newton-cg':
if daal_ready:
def make_ncg_funcs(f,
value=False,
gradient=False,
hessian=False):
daal_penaltyL2 = 0.5 / C / C_daal_multiplier
_obj_, X_, y_, n_samples = daal_extra_args_func(
classes.size,
w0,
X,
target,
0.,
daal_penaltyL2,
fit_intercept,
value=value,
gradient=gradient,
hessian=hessian)
_func_ = lambda x, *args: f(x, _obj_, *args)
return _func_, (X_, y_, n_samples, daal_penaltyL2)
loss_func, extra_args = make_ncg_funcs(func, value=True)
grad_func, _ = make_ncg_funcs(grad, gradient=True)
grad_hess_func, _ = make_ncg_funcs(hess, gradient=True)
w0, n_iter_i = newton_cg(grad_hess_func,
loss_func,
grad_func,
w0,
args=extra_args,
maxiter=max_iter,
tol=tol)
else:
args = (X, target, 1. / C, sample_weight)
w0, n_iter_i = newton_cg(hess,
func,
grad,
w0,
args=args,
maxiter=max_iter,
tol=tol)
elif solver == 'liblinear':
coef_, intercept_, n_iter_i, = _fit_liblinear(
X,
target,
C,
fit_intercept,
intercept_scaling,
None,
penalty,
dual,
verbose,
max_iter,
tol,
random_state,
sample_weight=sample_weight)
if fit_intercept:
w0 = np.concatenate([coef_.ravel(), intercept_])
else:
w0 = coef_.ravel()
elif solver in ['sag', 'saga']:
if multi_class == 'multinomial':
target = target.astype(X.dtype, copy=False)
loss = 'multinomial'
else:
loss = 'log'
# alpha is for L2-norm, beta is for L1-norm
if penalty == 'l1':
alpha = 0.
beta = 1. / C
elif penalty == 'l2':
alpha = 1. / C
beta = 0.
else: # Elastic-Net penalty
alpha = (1. / C) * (1 - l1_ratio)
beta = (1. / C) * l1_ratio
w0, n_iter_i, warm_start_sag = sag_solver(
X,
target,
sample_weight,
loss,
alpha,
beta,
max_iter,
tol,
verbose,
random_state,
False,
max_squared_sum,
warm_start_sag,
is_saga=(solver == 'saga'))
else:
raise ValueError("solver must be one of {'liblinear', 'lbfgs', "
"'newton-cg', 'sag'}, got '%s' instead" % solver)
if multi_class == 'multinomial':
if daal_ready:
if classes.size == 2:
multi_w0 = w0[np.newaxis, :]
else:
multi_w0 = np.reshape(w0, (classes.size, -1))
else:
n_classes = max(2, classes.size)
multi_w0 = np.reshape(w0, (n_classes, -1))
if n_classes == 2:
multi_w0 = multi_w0[1][np.newaxis, :]
coefs.append(np.require(multi_w0, requirements='O'))
else:
coefs.append(np.require(w0, requirements='O'))
n_iter[i] = n_iter_i
if daal_ready:
if fit_intercept:
for i, ci in enumerate(coefs):
coefs[i] = np.roll(ci, -1, -1)
else:
for i, ci in enumerate(coefs):
coefs[i] = np.delete(ci, 0, axis=-1)
return np.array(coefs), np.array(Cs), n_iter
def fit(self,X,y,sample_weight=None):
if not isinstance(self.C, numbers.Number) or self.C < 0:
raise ValueError("Penalty term must be positive; got (C=%r)"
% self.C)
if not isinstance(self.max_iter, numbers.Number) or self.max_iter < 0:
raise ValueError("Maximum number of iteration must be positive;"
" got (max_iter=%r)" % self.max_iter)
if not isinstance(self.tol, numbers.Number) or self.tol < 0:
raise ValueError("Tolerance for stopping criteria must be "
"positive; got (tol=%r)" % self.tol)
solver = _check_solver(self.solver, self.penalty, self.dual)
if solver in ['newton-cg']:
_dtype = [np.float64, np.float32]
else:
_dtype = np.float64
X, y = check_X_y(X, y, accept_sparse='csr', dtype=_dtype, order="C",
accept_large_sparse=solver != 'liblinear')
self.X_fit_ = X
X_k = self._get_kernel(X)
check_classification_targets(y)
self.classes_ = np.unique(y)
n_samples, n_features = X_k.shape
multi_class = _check_multi_class(self.multi_class, solver,
len(self.classes_))
if solver == 'liblinear':
if effective_n_jobs(self.n_jobs) != 1:
warnings.warn("'n_jobs' > 1 does not have any effect when"
" 'solver' is set to 'liblinear'. Got 'n_jobs'"
" = {}.".format(effective_n_jobs(self.n_jobs)))
self.coef_, self.intercept_, n_iter_ = _fit_liblinear(
X_k, y, self.C, self.fit_intercept, self.intercept_scaling,
self.class_weight, self.penalty, self.dual, self.verbose,
self.max_iter, self.tol, self.random_state,
sample_weight=sample_weight)
self.n_iter_ = np.array([n_iter_])
return self
if solver in ['sag', 'saga']:
max_squared_sum = row_norms(X_k, squared=True).max()
else:
max_squared_sum = None
n_classes = len(self.classes_)
classes_ = self.classes_
if n_classes < 2:
raise ValueError("This solver needs samples of at least 2 classes"
" in the data, but the data contains only one"
" class: %r" % classes_[0])
if len(self.classes_) == 2:
n_classes = 1
classes_ = classes_[1:]
if self.warm_start:
warm_start_coef = getattr(self, 'coef_', None)
else:
warm_start_coef = None
if warm_start_coef is not None and self.fit_intercept:
warm_start_coef = np.append(warm_start_coef,
self.intercept_[:, np.newaxis],
axis=1)
self.coef_ = list()
self.intercept_ = np.zeros(n_classes)
# Hack so that we iterate only once for the multinomial case.
if multi_class == 'multinomial':
classes_ = [None]
warm_start_coef = [warm_start_coef]
if warm_start_coef is None:
warm_start_coef = [None] * n_classes
path_func = delayed(logistic_regression_path)
# The SAG solver releases the GIL so it's more efficient to use
# threads for this solver.
if solver in ['sag', 'saga']:
prefer = 'threads'
else:
prefer = 'processes'
fold_coefs_ = Parallel(n_jobs=self.n_jobs, verbose=self.verbose,
**_joblib_parallel_args(prefer=prefer))(
path_func(X_k, y, pos_class=class_, Cs=[self.C],
fit_intercept=self.fit_intercept, tol=self.tol,
verbose=self.verbose, solver=solver,
multi_class=multi_class, max_iter=self.max_iter,
class_weight=self.class_weight, check_input=False,
random_state=self.random_state, coef=warm_start_coef_,
penalty=self.penalty,
max_squared_sum=max_squared_sum,
sample_weight=sample_weight)
for class_, warm_start_coef_ in zip(classes_, warm_start_coef))
fold_coefs_, _, n_iter_ = zip(*fold_coefs_)
self.n_iter_ = np.asarray(n_iter_, dtype=np.int32)[:, 0]
if multi_class == 'multinomial':
self.coef_ = fold_coefs_[0][0]
else:
self.coef_ = np.asarray(fold_coefs_)
self.coef_ = self.coef_.reshape(n_classes, n_features +
int(self.fit_intercept))
if self.fit_intercept:
self.intercept_ = self.coef_[:, -1]
self.coef_ = self.coef_[:, :-1]
return self
def _logistic_regression_path(X,
y,
Cs=48,
fit_intercept=True,
max_iter=100,
tol=1e-4,
verbose=0,
coef=None,
class_weight=None,
penalty='l2',
multi_class='auto',
check_input=True,
sample_weight=None,
l1_ratio=None,
coef_mask=None):
"""Compute a Logistic Regression model for a list of regularization
parameters.
Parameters
----------
X : array-like or sparse matrix, shape (n_samples, n_features)
Input data.
y : array-like, shape (n_samples,) or (n_samples, n_targets)
Input data, target values.
Cs : int | array-like, shape (n_cs,)
List of values for the regularization parameter or integer specifying
the number of regularization parameters that should be used. In this
case, the parameters will be chosen in a logarithmic scale between
1e-4 and 1e4.
fit_intercept : bool
Whether to fit an intercept for the model. In this case the shape of
the returned array is (n_cs, n_features + 1).
max_iter : int
Maximum number of iterations for the solver.
tol : float
Stopping criterion. For the newton-cg and lbfgs solvers, the iteration
will stop when ``max{|g_i | i = 1, ..., n} <= tol``
where ``g_i`` is the i-th component of the gradient.
verbose : int
For the liblinear and lbfgs solvers set verbose to any positive
number for verbosity.
coef : array-like, shape (n_features,), default None
Initialization value for coefficients of logistic regression.
Useless for liblinear solver.
class_weight : dict or 'balanced', optional
Weights associated with classes in the form ``{class_label: weight}``.
If not given, all classes are supposed to have weight one.
The "balanced" mode uses the values of y to automatically adjust
weights inversely proportional to class frequencies in the input data
as ``n_samples / (n_classes * np.bincount(y))``.
Note that these weights will be multiplied with sample_weight (passed
through the fit method) if sample_weight is specified.
multi_class : str, {'multinomial', 'auto'}, default: 'auto'
For 'multinomial' the loss minimised is the multinomial loss fit
across the entire probability distribution, *even when the data is
binary*. 'auto' selects binary if the data is binary
and otherwise selects 'multinomial'.
check_input : bool, default True
If False, the input arrays X and y will not be checked.
sample_weight : array-like, shape(n_samples,) optional
Array of weights that are assigned to individual samples.
If not provided, then each sample is given unit weight.
coef_mask : array-like, shape (n_features), (n_classes, n_features) optional
Masking array for coef.
Returns
-------
coefs : ndarray, shape (n_cs, n_features) or (n_cs, n_features + 1)
List of coefficients for the Logistic Regression model. If
fit_intercept is set to True then the second dimension will be
n_features + 1, where the last item represents the intercept. For
``multiclass='multinomial'``, the shape is (n_classes, n_cs,
n_features) or (n_classes, n_cs, n_features + 1).
Cs : ndarray
Grid of Cs used for cross-validation.
n_iter : array, shape (n_cs,)
Actual number of iteration for each Cs.
"""
solver = 'lbfgs'
if isinstance(Cs, numbers.Integral):
Cs = np.logspace(-4, 4, Cs)
# Preprocessing.
if check_input:
X = check_array(X,
accept_sparse='csr',
dtype=np.float64,
accept_large_sparse=True)
y = check_array(y, ensure_2d=False, dtype=None)
check_consistent_length(X, y)
_, n_features = X.shape
classes = np.unique(y)
multi_class = _check_multi_class(multi_class, solver, len(classes))
# If sample weights exist, convert them to array (support for lists)
# and check length
# Otherwise set them to 1 for all examples
if sample_weight is not None:
sample_weight = np.array(sample_weight, dtype=X.dtype, order='C')
check_consistent_length(y, sample_weight)
else:
sample_weight = np.ones(X.shape[0], dtype=X.dtype)
# If class_weights is a dict (provided by the user), the weights
# are assigned to the original labels. If it is "balanced", then
# the class_weights are assigned after masking the labels with a OvR.
le = LabelEncoder()
if isinstance(class_weight, dict) or multi_class == 'multinomial':
class_weight_ = compute_class_weight(class_weight, classes, y)
sample_weight *= class_weight_[le.fit_transform(y)]
# For doing a ovr, we need to mask the labels first. for the
# multinomial case this is not necessary.
if multi_class == 'ovr':
coef_size = n_features
w0 = np.zeros(n_features + int(fit_intercept), dtype=X.dtype)
mask_classes = np.array([-1, 1])
mask = (y == 1)
y_bin = np.ones(y.shape, dtype=X.dtype)
y_bin[~mask] = -1.
# for compute_class_weight
if class_weight == "balanced":
class_weight_ = compute_class_weight(class_weight, mask_classes,
y_bin)
sample_weight *= class_weight_[le.fit_transform(y_bin)]
else:
coef_size = classes.size * n_features
lbin = OneHotEncoder(categories=[range(classes.size)], sparse=False)
Y_multi = lbin.fit_transform(y[:, np.newaxis])
if Y_multi.shape[1] == 1:
Y_multi = np.hstack([1 - Y_multi, Y_multi])
w0 = np.zeros((classes.size, n_features + int(fit_intercept)),
dtype=X.dtype)
w0[:, -1] = LogisticInterceptFitterNoFeatures(y,
classes.size).intercept_
if coef is not None:
# it must work both giving the bias term and not
if multi_class == 'ovr':
if coef.size not in (n_features, w0.size):
raise ValueError(
'Initialization coef is of shape %d, expected shape '
'%d or %d' % (coef.size, n_features, w0.size))
w0[:coef.size] = coef
else:
w0[:, :coef.shape[1]] = coef
# Mask initial array
if coef_mask is not None:
if multi_class == 'ovr':
w0[:n_features] *= coef_mask
else:
w0[:, :n_features] *= coef_mask
if multi_class == 'multinomial':
# fmin_l_bfgs_b and newton-cg accepts only ravelled parameters.
target = Y_multi
if penalty == 'l2':
w0 = w0.ravel()
def func(x, *args):
return _multinomial_loss_grad(x, *args)[0:2]
else:
w0 = w0.T.ravel().copy()
def inner_func(x, *args):
return _multinomial_loss_grad(x, *args)[0:2]
def func(x, g, *args):
x = x.reshape(-1, classes.size).T.ravel()
loss, grad = inner_func(x, *args)
grad = grad.reshape(classes.size, -1).T.ravel()
g[:] = grad
return loss
else:
target = y_bin
if penalty == 'l2':
func = _logistic_loss_and_grad
else:
def func(x, g, *args):
loss, grad = _logistic_loss_and_grad(x, *args)
g[:] = grad
return loss
coefs = list()
n_iter = np.zeros(len(Cs), dtype=np.int32)
for i, C in enumerate(Cs):
iprint = [-1, 50, 1, 100, 101][np.searchsorted(np.array([0, 1, 2, 3]),
verbose)]
if penalty == 'l2':
w0, loss, info = optimize.fmin_l_bfgs_b(func,
w0,
fprime=None,
args=(X, target, 1. / C,
coef_mask,
sample_weight),
iprint=iprint,
pgtol=tol,
maxiter=max_iter)
else:
zeros_seen = [0]
def zero_coef(x, *args):
if multi_class == 'multinomial':
x = x.reshape(-1, classes.size)[:-1]
else:
x = x[:-1]
now_zeros = np.array_equiv(x, 0.)
if now_zeros:
zeros_seen[0] += 1
else:
zeros_seen[0] = 0
if zeros_seen[0] > 1:
return -2048
try:
w0 = fmin_lbfgs(func,
w0,
orthantwise_c=1. / C,
args=(X, target, 0., coef_mask, sample_weight),
max_iterations=max_iter,
epsilon=tol,
orthantwise_end=coef_size,
progress=zero_coef)
except AllZeroLBFGSError:
w0 *= 0.
info = None
if info is not None and info["warnflag"] == 1:
warnings.warn(
"lbfgs failed to converge. Increase the number "
"of iterations.", ConvergenceWarning)
# In scipy <= 1.0.0, nit may exceed maxiter.
# See https://github.com/scipy/scipy/issues/7854.
if info is None:
n_iter_i = -1
else:
n_iter_i = min(info['nit'], max_iter)
if multi_class == 'multinomial':
n_classes = max(2, classes.size)
if penalty == 'l2':
multi_w0 = np.reshape(w0, (n_classes, -1))
else:
multi_w0 = np.reshape(w0, (-1, n_classes)).T
if coef_mask is not None:
multi_w0[:, :n_features] *= coef_mask
coefs.append(multi_w0.copy())
else:
if coef_mask is not None:
w0[:n_features] *= coef_mask
coefs.append(w0.copy())
n_iter[i] = n_iter_i
return np.array(coefs), np.array(Cs), n_iter
def fit(self, X, y, sample_weight=None, coef_mask=None):
"""Fit the model according to the given training data.
Parameters
----------
X : {array-like, sparse matrix}, shape (n_samples, n_features)
Training vector, where n_samples is the number of samples and
n_features is the number of features.
y : array-like, shape (n_samples,)
Target vector relative to X.
sample_weight : array-like, shape (n_samples,) optional
Array of weights that are assigned to individual samples.
If not provided, then each sample is given unit weight.
coef_mask : array-like, shape (n_features), (n_classes, n_features)
optional
Masking array for coef.
Returns
-------
self : object
"""
solver = 'lbfgs'
if not isinstance(self.C, numbers.Number) or self.C < 0:
raise ValueError("Penalty term must be positive; got (C=%r)" %
self.C)
if not isinstance(self.max_iter, numbers.Number) or self.max_iter < 0:
raise ValueError("Maximum number of iteration must be positive;"
" got (max_iter=%r)" % self.max_iter)
if not isinstance(self.tol, numbers.Number) or self.tol < 0:
raise ValueError("Tolerance for stopping criteria must be "
"positive; got (tol=%r)" % self.tol)
_dtype = np.float64
X, y = self._pre_fit(X, y)
X, y = check_X_y(X,
y,
accept_sparse='csr',
dtype=_dtype,
order="C",
accept_large_sparse=True)
check_classification_targets(y)
self.classes_ = np.unique(y)
n_samples, n_features = X.shape
multi_class = _check_multi_class(self.multi_class, solver,
len(self.classes_))
n_classes = len(self.classes_)
classes_ = self.classes_
if n_classes < 2:
raise ValueError("This solver needs samples of at least 2 classes"
" in the data, but the data contains only one"
" class: %r" % classes_[0])
if len(self.classes_) == 2 and multi_class == 'ovr':
n_classes = 1
classes_ = classes_[1:]
if multi_class == 'multinomial' and coef_mask is not None:
coef_mask = coef_mask.reshape(n_classes, -1)
if self.warm_start:
warm_start_coef = getattr(self, 'coef_', None)
else:
warm_start_coef = None
if warm_start_coef is not None and self.fit_intercept:
warm_start_coef = np.append(warm_start_coef,
self.intercept_[:, np.newaxis],
axis=1)
self.intercept_ = np.zeros(n_classes)
fold_coefs_ = _logistic_regression_path(
X,
y,
Cs=[self.C],
fit_intercept=self.fit_intercept,
tol=self.tol,
verbose=self.verbose,
multi_class=multi_class,
max_iter=self.max_iter,
class_weight=self.class_weight,
penalty=self.penalty,
check_input=False,
coef=warm_start_coef,
sample_weight=sample_weight,
coef_mask=coef_mask)
fold_coefs_, _, self.n_iter_ = fold_coefs_
self.coef_ = np.asarray(fold_coefs_)
self.coef_ = self.coef_.reshape(n_classes,
n_features + int(self.fit_intercept))
if self.fit_intercept:
self.intercept_ = self.coef_[:, -1]
self.coef_ = self.coef_[:, :-1]
self._post_fit(X, y)
return self
def _logistic_regression_path(X,
y,
pos_class=None,
Cs=10,
fit_intercept=True,
max_iter=100,
tol=1e-4,
verbose=0,
coef=None,
class_weight=None,
penalty='l2',
multi_class='warn',
random_state=None,
check_input=True,
sample_weight=None,
l1_ratio=None,
coef_mask=None):
"""Compute a Logistic Regression model for a list of regularization
parameters.
This is an implementation that uses the result of the previous model
to speed up computations along the set of solutions, making it faster
than sequentially calling LogisticRegression for the different parameters.
Note that there will be no speedup with liblinear solver, since it does
not handle warm-starting.
Read more in the :ref:`User Guide <logistic_regression>`.
Parameters
----------
X : array-like or sparse matrix, shape (n_samples, n_features)
Input data.
y : array-like, shape (n_samples,) or (n_samples, n_targets)
Input data, target values.
pos_class : int, None
The class with respect to which we perform a one-vs-all fit.
If None, then it is assumed that the given problem is binary.
Cs : int | array-like, shape (n_cs,)
List of values for the regularization parameter or integer specifying
the number of regularization parameters that should be used. In this
case, the parameters will be chosen in a logarithmic scale between
1e-4 and 1e4.
fit_intercept : bool
Whether to fit an intercept for the model. In this case the shape of
the returned array is (n_cs, n_features + 1).
max_iter : int
Maximum number of iterations for the solver.
tol : float
Stopping criterion. For the newton-cg and lbfgs solvers, the iteration
will stop when ``max{|g_i | i = 1, ..., n} <= tol``
where ``g_i`` is the i-th component of the gradient.
verbose : int
For the liblinear and lbfgs solvers set verbose to any positive
number for verbosity.
coef : array-like, shape (n_features,), default None
Initialization value for coefficients of logistic regression.
Useless for liblinear solver.
class_weight : dict or 'balanced', optional
Weights associated with classes in the form ``{class_label: weight}``.
If not given, all classes are supposed to have weight one.
The "balanced" mode uses the values of y to automatically adjust
weights inversely proportional to class frequencies in the input data
as ``n_samples / (n_classes * np.bincount(y))``.
Note that these weights will be multiplied with sample_weight (passed
through the fit method) if sample_weight is specified.
multi_class : str, {'ovr', 'multinomial', 'auto'}, default: 'ovr'
If the option chosen is 'ovr', then a binary problem is fit for each
label. For 'multinomial' the loss minimised is the multinomial loss fit
across the entire probability distribution, *even when the data is
binary*. 'multinomial' is unavailable when solver='liblinear'.
'auto' selects 'ovr' if the data is binary, or if solver='liblinear',
and otherwise selects 'multinomial'.
.. versionadded:: 0.18
Stochastic Average Gradient descent solver for 'multinomial' case.
.. versionchanged:: 0.20
Default will change from 'ovr' to 'auto' in 0.22.
random_state : int, RandomState instance or None, optional, default None
The seed of the pseudo random number generator to use when shuffling
the data. If int, random_state is the seed used by the random number
generator; If RandomState instance, random_state is the random number
generator; If None, the random number generator is the RandomState
instance used by `np.random`. Used when ``solver`` == 'sag' or
'liblinear'.
check_input : bool, default True
If False, the input arrays X and y will not be checked.
sample_weight : array-like, shape(n_samples,) optional
Array of weights that are assigned to individual samples.
If not provided, then each sample is given unit weight.
coef_mask : array-like, shape (n_features), (n_classes, n_features) optional
Masking array for coef.
Returns
-------
coefs : ndarray, shape (n_cs, n_features) or (n_cs, n_features + 1)
List of coefficients for the Logistic Regression model. If
fit_intercept is set to True then the second dimension will be
n_features + 1, where the last item represents the intercept. For
``multiclass='multinomial'``, the shape is (n_classes, n_cs,
n_features) or (n_classes, n_cs, n_features + 1).
Cs : ndarray
Grid of Cs used for cross-validation.
n_iter : array, shape (n_cs,)
Actual number of iteration for each Cs.
Notes
-----
You might get slightly different results with the solver liblinear than
with the others since this uses LIBLINEAR which penalizes the intercept.
.. versionchanged:: 0.19
The "copy" parameter was removed.
"""
solver = 'lbfgs'
if isinstance(Cs, numbers.Integral):
Cs = np.logspace(-4, 4, Cs)
# Preprocessing.
if check_input:
X = check_array(X,
accept_sparse='csr',
dtype=np.float64,
accept_large_sparse=True)
y = check_array(y, ensure_2d=False, dtype=None)
check_consistent_length(X, y)
_, n_features = X.shape
classes = np.unique(y)
random_state = check_random_state(random_state)
multi_class = _check_multi_class(multi_class, solver, len(classes))
if pos_class is None and multi_class != 'multinomial':
if (classes.size > 2):
raise ValueError('To fit OvR, use the pos_class argument')
# np.unique(y) gives labels in sorted order.
pos_class = classes[1]
# If sample weights exist, convert them to array (support for lists)
# and check length
# Otherwise set them to 1 for all examples
if sample_weight is not None:
sample_weight = np.array(sample_weight, dtype=X.dtype, order='C')
check_consistent_length(y, sample_weight)
else:
sample_weight = np.ones(X.shape[0], dtype=X.dtype)
# If class_weights is a dict (provided by the user), the weights
# are assigned to the original labels. If it is "balanced", then
# the class_weights are assigned after masking the labels with a OvR.
le = LabelEncoder()
if isinstance(class_weight, dict) or multi_class == 'multinomial':
class_weight_ = compute_class_weight(class_weight, classes, y)
sample_weight *= class_weight_[le.fit_transform(y)]
# For doing a ovr, we need to mask the labels first. for the
# multinomial case this is not necessary.
if multi_class == 'ovr':
coef_size = n_features
w0 = np.zeros(n_features + int(fit_intercept), dtype=X.dtype)
mask_classes = np.array([-1, 1])
mask = (y == pos_class)
y_bin = np.ones(y.shape, dtype=X.dtype)
y_bin[~mask] = -1.
# for compute_class_weight
if class_weight == "balanced":
class_weight_ = compute_class_weight(class_weight, mask_classes,
y_bin)
sample_weight *= class_weight_[le.fit_transform(y_bin)]
else:
coef_size = classes.size * n_features
lbin = LabelBinarizer()
Y_multi = lbin.fit_transform(y)
if Y_multi.shape[1] == 1:
Y_multi = np.hstack([1 - Y_multi, Y_multi])
w0 = np.zeros((classes.size, n_features + int(fit_intercept)),
dtype=X.dtype)
if coef is not None:
# it must work both giving the bias term and not
if multi_class == 'ovr':
if coef.size not in (n_features, w0.size):
raise ValueError(
'Initialization coef is of shape %d, expected shape '
'%d or %d' % (coef.size, n_features, w0.size))
w0[:coef.size] = coef
else:
# For binary problems coef.shape[0] should be 1, otherwise it
# should be classes.size.
n_classes = classes.size
if n_classes == 2:
n_classes = 1
if (coef.shape[0] != n_classes
or coef.shape[1] not in (n_features, n_features + 1)):
raise ValueError(
'Initialization coef is of shape (%d, %d), expected '
'shape (%d, %d) or (%d, %d)' %
(coef.shape[0], coef.shape[1], classes.size, n_features,
classes.size, n_features + 1))
if n_classes == 1:
w0[0, :coef.shape[1]] = -coef
w0[1, :coef.shape[1]] = coef
else:
w0[:, :coef.shape[1]] = coef
# Mask initial array
if coef_mask is not None:
if multi_class == 'ovr':
w0[:n_features] *= coef_mask
else:
w0[:, :n_features] *= coef_mask
if multi_class == 'multinomial':
# fmin_l_bfgs_b and newton-cg accepts only ravelled parameters.
target = Y_multi
if penalty == 'l2':
w0 = w0.ravel()
def func(x, *args):
return _multinomial_loss_grad(x, *args)[0:2]
else:
w0 = w0.T.ravel().copy()
def inner_func(x, *args):
return _multinomial_loss_grad(x, *args)[0:2]
def func(x, g, *args):
x = x.reshape(-1, classes.size).T.ravel()
loss, grad = inner_func(x, *args)
grad = grad.reshape(classes.size, -1).T.ravel()
g[:] = grad
return loss
else:
target = y_bin
if penalty == 'l2':
func = _logistic_loss_and_grad
else:
def func(x, g, *args):
loss, grad = _logistic_loss_and_grad(x, *args)
g[:] = grad
return loss
coefs = list()
n_iter = np.zeros(len(Cs), dtype=np.int32)
for i, C in enumerate(Cs):
iprint = [-1, 50, 1, 100, 101][np.searchsorted(np.array([0, 1, 2, 3]),
verbose)]
if penalty == 'l2':
w0, loss, info = optimize.fmin_l_bfgs_b(func,
w0,
fprime=None,
args=(X, target, 1. / C,
coef_mask,
sample_weight),
iprint=iprint,
pgtol=tol,
maxiter=max_iter)
else:
w0 = fmin_lbfgs(func,
w0,
orthantwise_c=1. / C,
args=(X, target, 0., coef_mask, sample_weight),
max_iterations=max_iter,
epsilon=tol,
orthantwise_end=coef_size)
info = None
if info is not None and info["warnflag"] == 1:
warnings.warn(
"lbfgs failed to converge. Increase the number "
"of iterations.", ConvergenceWarning)
# In scipy <= 1.0.0, nit may exceed maxiter.
# See https://github.com/scipy/scipy/issues/7854.
if info is None:
n_iter_i = -1
else:
n_iter_i = min(info['nit'], max_iter)
if multi_class == 'multinomial':
n_classes = max(2, classes.size)
if penalty == 'l2':
multi_w0 = np.reshape(w0, (n_classes, -1))
else:
multi_w0 = np.reshape(w0, (-1, n_classes)).T
if coef_mask is not None:
multi_w0[:, :n_features] *= coef_mask
if n_classes == 2:
multi_w0 = multi_w0[1][np.newaxis, :]
coefs.append(multi_w0.copy())
else:
if coef_mask is not None:
w0[:n_features] *= coef_mask
coefs.append(w0.copy())
n_iter[i] = n_iter_i
return np.array(coefs), np.array(Cs), n_iter
def fit(self, X, y, sample_weight=None, coef_mask=None):
"""Fit the model according to the given training data.
Parameters
----------
X : {array-like, sparse matrix}, shape (n_samples, n_features)
Training vector, where n_samples is the number of samples and
n_features is the number of features.
y : array-like, shape (n_samples,)
Target vector relative to X.
sample_weight : array-like, shape (n_samples,) optional
Array of weights that are assigned to individual samples.
If not provided, then each sample is given unit weight.
.. versionadded:: 0.17
*sample_weight* support to LogisticRegression.
coef_mask : array-like, shape (n_features), (n_classes, n_features)
optional
Masking array for coef.
Returns
-------
self : object
Notes
-----
"""
solver = 'lbfgs'
if not isinstance(self.C, numbers.Number) or self.C < 0:
raise ValueError("Penalty term must be positive; got (C=%r)" %
self.C)
if not isinstance(self.max_iter, numbers.Number) or self.max_iter < 0:
raise ValueError("Maximum number of iteration must be positive;"
" got (max_iter=%r)" % self.max_iter)
if not isinstance(self.tol, numbers.Number) or self.tol < 0:
raise ValueError("Tolerance for stopping criteria must be "
"positive; got (tol=%r)" % self.tol)
_dtype = np.float64
X, y = check_X_y(X,
y,
accept_sparse='csr',
dtype=_dtype,
order="C",
accept_large_sparse=True)
check_classification_targets(y)
self.classes_ = np.unique(y)
n_samples, n_features = X.shape
multi_class = _check_multi_class(self.multi_class, solver,
len(self.classes_))
n_classes = len(self.classes_)
if multi_class == 'multinomial' and coef_mask is not None:
coef_mask = coef_mask.reshape(n_classes, -1)
classes_ = self.classes_
if n_classes < 2:
raise ValueError("This solver needs samples of at least 2 classes"
" in the data, but the data contains only one"
" class: %r" % classes_[0])
if len(self.classes_) == 2:
n_classes = 1
classes_ = classes_[1:]
if self.warm_start:
warm_start_coef = getattr(self, 'coef_', None)
else:
warm_start_coef = None
if warm_start_coef is not None and self.fit_intercept:
warm_start_coef = np.append(warm_start_coef,
self.intercept_[:, np.newaxis],
axis=1)
self.coef_ = list()
self.intercept_ = np.zeros(n_classes)
# Hack so that we iterate only once for the multinomial case.
if multi_class == 'multinomial':
classes_ = [None]
warm_start_coef = [warm_start_coef]
if warm_start_coef is None:
warm_start_coef = [None] * n_classes
path_func = delayed(_logistic_regression_path)
# The SAG solver releases the GIL so it's more efficient to use
# threads for this solver.
prefer = 'processes'
fold_coefs_ = Parallel(
n_jobs=self.n_jobs,
verbose=self.verbose,
**_joblib_parallel_args(prefer=prefer))(
path_func(X,
y,
pos_class=class_,
Cs=[self.C],
fit_intercept=self.fit_intercept,
tol=self.tol,
verbose=self.verbose,
multi_class=multi_class,
max_iter=self.max_iter,
class_weight=self.class_weight,
penalty=self.penalty,
check_input=False,
random_state=self.random_state,
coef=warm_start_coef_,
sample_weight=sample_weight,
coef_mask=coef_mask)
for class_, warm_start_coef_ in zip(classes_, warm_start_coef))
fold_coefs_, _, n_iter_ = zip(*fold_coefs_)
self.n_iter_ = np.asarray(n_iter_, dtype=np.int32)[:, 0]
if multi_class == 'multinomial':
self.coef_ = fold_coefs_[0][0]
else:
self.coef_ = np.asarray(fold_coefs_)
self.coef_ = self.coef_.reshape(
n_classes, n_features + int(self.fit_intercept))
if self.fit_intercept:
self.intercept_ = self.coef_[:, -1]
self.coef_ = self.coef_[:, :-1]
return self
