def sgd_regression(name,
loss=None, #default - 'hinge'
penalty=None, #default - 'l2'
alpha=None, #default - 0.0001
l1_ratio=None, #default - 0.15, must be within [0, 1]
fit_intercept=None, #default - True
n_iter=None, #default - 5
shuffle=None, #default - False
random_state=None, #default - None
epsilon=None, #default - 0.1
learning_rate=None, #default - 'invscaling'
eta0=None, #default - 0.01
power_t=None, #default - 0.5
warm_start=False,
verbose=0,
):
def _name(msg):
return '%s.%s_%s' % (name, 'sgd', msg)
rval = scope.sklearn_SGDRegressor(
loss=hp.pchoice(
_name('loss'),
[ (0.25, 'squared_loss'),
(0.25, 'huber'),
(0.25, 'epsilon_insensitive'),
(0.25, 'squared_epsilon_insensitive') ] ) if loss is None else loss,
penalty=hp.pchoice(
_name('penalty'),
[ (0.40, 'l2'),
(0.35, 'l1'),
(0.25, 'elasticnet') ] ) if penalty is None else penalty,
alpha=hp.loguniform(
_name('alpha'),
np.log(1e-7),
np.log(1)) if alpha is None else alpha,
l1_ratio=hp.uniform(
_name('l1_ratio'),
0, 1 ) if l1_ratio is None else l1_ratio,
fit_intercept=hp.pchoice(
_name('fit_intercept'),
[ (0.8, True), (0.2, False) ]) if fit_intercept is None else fit_intercept,
epsilon=hp.loguniform(
_name('epsilon'),
np.log(1e-7),
np.log(1)) if epsilon is None else epsilon,
learning_rate='invscaling' if learning_rate is None else learning_rate,
eta0=hp.loguniform(
_name('eta0'),
np.log(1e-5),
np.log(1e-1)) if eta0 is None else eta0,
power_t=hp.uniform(
_name('power_t'),
0, 1) if power_t is None else power_t,
verbose=verbose,
random_state=random_state,
)
return rval
def sgd_regression(
name,
loss=None, #default - 'hinge'
penalty=None, #default - 'l2'
alpha=None, #default - 0.0001
l1_ratio=None, #default - 0.15, must be within [0, 1]
fit_intercept=None, #default - True
n_iter=None, #default - 5
shuffle=None, #default - False
random_state=None, #default - None
epsilon=None, #default - 0.1
learning_rate=None, #default - 'invscaling'
eta0=None, #default - 0.01
power_t=None, #default - 0.5
warm_start=False,
verbose=0,
):
def _name(msg):
return '%s.%s_%s' % (name, 'sgd', msg)
rval = scope.sklearn_SGDRegressor(
loss=hp.pchoice(_name('loss'), [(0.25, 'squared_loss'),
(0.25, 'huber'),
(0.25, 'epsilon_insensitive'),
(0.25, 'squared_epsilon_insensitive')])
if loss is None else loss,
penalty=hp.pchoice(_name('penalty'), [(0.40, 'l2'), (0.35, 'l1'),
(0.25, 'elasticnet')])
if penalty is None else penalty,
alpha=hp.loguniform(_name('alpha'), np.log(1e-7), np.log(1))
if alpha is None else alpha,
l1_ratio=hp.uniform(_name('l1_ratio'), 0, 1)
if l1_ratio is None else l1_ratio,
fit_intercept=hp.pchoice(_name('fit_intercept'), [(0.8, True),
(0.2, False)])
if fit_intercept is None else fit_intercept,
epsilon=hp.loguniform(_name('epsilon'), np.log(1e-7), np.log(1))
if epsilon is None else epsilon,
learning_rate='invscaling' if learning_rate is None else learning_rate,
eta0=hp.loguniform(_name('eta0'), np.log(1e-5), np.log(1e-1))
if eta0 is None else eta0,
power_t=hp.uniform(_name('power_t'), 0, 1)
if power_t is None else power_t,
verbose=verbose,
random_state=random_state,
)
return rval
def sgd_regression(name,
loss=None, # default - 'squared_loss'
penalty=None, # default - 'l2'
alpha=None, # default - 0.0001
l1_ratio=None, # default - 0.15, must be within [0, 1]
fit_intercept=True, # default - True
n_iter=5, # default - 5
shuffle=None, # default - False
random_state=None, # default - None
epsilon=None, # default - 0.1
learning_rate=None, # default - 'invscaling'
eta0=None, # default - 0.01
power_t=None, # default - 0.5
warm_start=False,
verbose=False):
def _name(msg):
return '%s.%s_%s' % (name, 'sgdr', msg)
rval = scope.sklearn_SGDRegressor(
loss=hp.pchoice(_name('loss'), [
(0.25, 'squared_loss'),
(0.25, 'huber'),
(0.25, 'epsilon_insensitive'),
(0.25, 'squared_epsilon_insensitive')
]) if loss is None else loss,
penalty=_sgd_penalty(_name('penalty')) if penalty is None else penalty,
alpha=_sgd_alpha(_name('alpha')) if alpha is None else alpha,
l1_ratio=(_sgd_l1_ratio(_name('l1ratio'))
if l1_ratio is None else l1_ratio),
fit_intercept=fit_intercept,
n_iter=n_iter,
# For regression, use the SVM epsilon instead of the SGD one.
epsilon=_svm_epsilon(_name('epsilon')) if epsilon is None else epsilon,
learning_rate=(_sgdr_learning_rate(_name('learning_rate'))
if learning_rate is None else learning_rate),
eta0=_sgd_eta0(_name('eta0')) if eta0 is None else eta0,
power_t=_sgd_power_t(_name('power_t')) if power_t is None else power_t,
verbose=verbose,
random_state=_random_state(_name('rstate'), random_state),
)
return rval
def sgd_regression(name,
loss=None, # default - 'squared_loss'
penalty=None, # default - 'l2'
alpha=None, # default - 0.0001
l1_ratio=None, # default - 0.15, must be within [0, 1]
fit_intercept=True, # default - True
n_iter=5, # default - 5
shuffle=None, # default - False
random_state=None, # default - None
epsilon=None, # default - 0.1
learning_rate=None, # default - 'invscaling'
eta0=None, # default - 0.01
power_t=None, # default - 0.5
warm_start=False,
verbose=False):
def _name(msg):
return '%s.%s_%s' % (name, 'sgdr', msg)
rval = scope.sklearn_SGDRegressor(
loss=hp.pchoice(_name('loss'), [
(0.25, 'squared_loss'),
(0.25, 'huber'),
(0.25, 'epsilon_insensitive'),
(0.25, 'squared_epsilon_insensitive')
]) if loss is None else loss,
penalty=_sgd_penalty(_name('penalty')) if penalty is None else penalty,
alpha=_sgd_alpha(_name('alpha')) if alpha is None else alpha,
l1_ratio=(_sgd_l1_ratio(_name('l1ratio'))
if l1_ratio is None else l1_ratio),
fit_intercept=fit_intercept,
n_iter=n_iter,
# For regression, use the SVM epsilon instead of the SGD one.
epsilon=_svm_epsilon(_name('epsilon')) if epsilon is None else epsilon,
learning_rate=(_sgdr_learning_rate(_name('learning_rate'))
if learning_rate is None else learning_rate),
eta0=_sgd_eta0(_name('eta0')) if eta0 is None else eta0,
power_t=_sgd_power_t(_name('power_t')) if power_t is None else power_t,
verbose=verbose,
random_state=_random_state(_name('rstate'), random_state),
)
return rval
