def _get_n_jobs(self, logger, **kwargs):
try:
if config.fixed_num_folds == 0:
n_jobs = max(1, int(int(max_threads() / min(config.num_folds, kwargs['max_workers']))))
else:
n_jobs = max(1, int(int(max_threads() / min(config.fixed_num_folds, config.num_folds, kwargs['max_workers']))))
except KeyError:
loggerinfo(logger, "Prophet No Max Worker in kwargs. Set n_jobs to 1")
n_jobs = 1
return n_jobs
def _get_n_jobs(logger, **kwargs):
if 'n_jobs_prophet' in config.recipe_dict:
return min(config.recipe_dict['n_jobs_prophet'], max_threads())
try:
if config.fixed_num_folds <= 0:
n_jobs = max(1, int(int(max_threads() / min(config.num_folds, kwargs['max_workers']))))
else:
n_jobs = max(1, int(
int(max_threads() / min(config.fixed_num_folds, config.num_folds, kwargs['max_workers']))))
except KeyError:
loggerinfo(logger, "Prophet No Max Worker in kwargs. Set n_jobs to 1")
n_jobs = 1
return n_jobs if n_jobs > 1 else 1
def update_n_jobs(self, X_shape, **kwargs):
if 'IS_SCORER' in kwargs:
# i.e. don't listen to fit-time n_jobs for predict if scoring
parent_max_workers = max(
1,
kwargs.get('parent_max_workers', kwargs.get('max_workers', 1))
or 1)
new_optimal = optimal_nthreads_model(
X_shape, max_workers=parent_max_workers)
if config.hard_asserts:
# then in testing situation, maybe multiple experiments, so drop down
self.params['n_jobs'] = self.params_base['n_jobs'] = min(
self.params_base['n_jobs'], new_optimal)
else:
self.params['n_jobs'] = self.params_base[
'n_jobs'] = new_optimal
else:
# still doing experiment, shared resources assumed, re-use self.params_base['n_jobs'] from fit
pass
if config.max_fit_cores_override > 0:
self.params_base['n_jobs'] = config.max_fit_cores_override
# self.params_base['n_jobs'] can change for predict
n_jobs = min(max_threads(), get_num_threads(),
self.params_base['n_jobs'])
return n_jobs
def set_default_params(self,
accuracy=None,
time_tolerance=None,
interpretability=None,
**kwargs):
self.params = dict(
random_state=kwargs.get("random_state", 24),
eta=0.1,
max_depth=12,
min_child_weight=2.0,
reg_lambda=1.0,
colsample_bytree=0.8,
subsample=1.0,
mu=0.1,
reg_alpha=0,
n_jobs=self.params_base.get('n_jobs', max_threads()),
)
def mutate_params(self, accuracy=10, **kwargs):
if accuracy > 8:
eta = [0.5, 0.1, 0.05, 0.01]
max_depth = list(range(4, 21))
min_child_weight = [0.1, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0, 32.0]
reg_lambda = [0.0, 0.1, 1.0, 2.0, 5.0, 8.0, 10.0, 20.0]
reg_alpha = [0.0, 0.1, 1.0, 5.0, 10.0]
colsample_bytree = [0.1 * ii for ii in range(1, 11)]
subsample = [0.5, 0.8, 0.9, 1.0]
mu = [0.05 * ii for ii in range(1, 14)]
elif accuracy >= 5:
eta = [0.5, 0.1, 0.05]
max_depth = list(range(4, 21, 2))
min_child_weight = [0.1, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0, 32.0]
reg_lambda = [0.0, 0.1, 1.0, 2.0, 5.0, 8.0, 10.0, 20.0]
reg_alpha = [0.0, 0.1, 1.0]
colsample_bytree = [0.1 * ii for ii in range(2, 11, 2)]
subsample = [1.0]
mu = [0.05 * ii for ii in range(1, 14)]
else:
eta = [0.1]
max_depth = list(range(4, 21, 2))
min_child_weight = [0.1, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0, 32.0]
reg_lambda = [0.0, 0.1, 1.0, 5.0, 10.0]
reg_alpha = [0.0]
colsample_bytree = [0.1 * ii for ii in range(2, 11, 2)]
subsample = [1.0]
mu = [0.05 * ii for ii in range(1, 14)]
self.params["eta"] = np.random.choice(eta)
self.params["max_depth"] = np.random.choice(max_depth)
self.params["min_child_weight"] = np.random.choice(min_child_weight)
self.params["reg_lambda"] = np.random.choice(reg_lambda)
self.params["reg_alpha"] = np.random.choice(reg_alpha)
self.params["colsample_bytree"] = np.random.choice(colsample_bytree)
self.params["subsample"] = np.random.choice(subsample)
self.params["mu"] = np.random.choice(mu)
self.params["n_jobs"] = self.params_base.get('n_jobs', max_threads())
