def _fit_stages(self,
X,
y,
raw_predictions,
sample_weight,
random_state,
begin_at_stage=0,
monitor=None,
X_idx_sorted=None):
"""Iteratively fits the stages.
For each stage it computes the progress (OOB, train score)
and delegates to ``_fit_stage``.
Returns the number of stages fit; might differ from ``n_estimators``
due to early stopping.
"""
n_samples = X.shape[0]
do_oob = self.subsample < 1.0
sample_mask = numpy.ones((n_samples, ), dtype=numpy.bool)
n_inbag = max(1, int(self.subsample * n_samples))
loss_ = self.loss_
if self.verbose:
verbose_reporter = VerboseReporter(verbose=self.verbose)
verbose_reporter.init(self, begin_at_stage)
X_csc = csc_matrix(X) if issparse(X) else None
X_csr = csr_matrix(X) if issparse(X) else None
if self.dropout_rate > 0.:
scale = numpy.ones(self.n_estimators, dtype=float)
else:
scale = None
# perform boosting iterations
i = begin_at_stage
for i in range(begin_at_stage, self.n_estimators):
# subsampling
if do_oob:
sample_mask = _random_sample_mask(n_samples, n_inbag,
random_state)
# OOB score before adding this stage
y_oob_sample = y[~sample_mask]
old_oob_score = loss_(y_oob_sample,
raw_predictions[~sample_mask],
sample_weight[~sample_mask])
# fit next stage of trees
raw_predictions = self._fit_stage(i, X, y, raw_predictions,
sample_weight, sample_mask,
random_state, scale,
X_idx_sorted, X_csc, X_csr)
# track deviance (= loss)
if do_oob:
self.train_score_[i] = loss_(y[sample_mask],
raw_predictions[sample_mask],
sample_weight[sample_mask])
self.oob_improvement_[i] = (
old_oob_score -
loss_(y_oob_sample, raw_predictions[~sample_mask],
sample_weight[~sample_mask]))
else:
# no need to fancy index w/ no subsampling
self.train_score_[i] = loss_(y, raw_predictions, sample_weight)
if self.verbose > 0:
verbose_reporter.update(i, self)
if monitor is not None:
early_stopping = monitor(i, self, locals())
if early_stopping:
break
if self.dropout_rate > 0.:
self.scale_ = scale
return i + 1
def _fit(self, X, event, time, sample_weight, random_state): # noqa: C901
n_samples = X.shape[0]
# account for intercept
Xi = numpy.column_stack((numpy.ones(n_samples), X))
y = numpy.fromiter(zip(event, time),
dtype=[('event', numpy.bool),
('time', numpy.float64)])
y_pred = numpy.zeros(n_samples)
do_oob = self.subsample < 1.0
if do_oob:
n_inbag = max(1, int(self.subsample * n_samples))
do_dropout = self.dropout_rate > 0
if do_dropout:
scale = numpy.ones(int(self.n_estimators), dtype=float)
if self.verbose:
verbose_reporter = VerboseReporter(verbose=self.verbose)
verbose_reporter.init(self, 0)
for num_iter in range(int(self.n_estimators)):
if do_oob:
sample_mask = _random_sample_mask(n_samples, n_inbag,
random_state)
subsample_weight = sample_weight * sample_mask.astype(
numpy.float64)
# OOB score before adding this stage
old_oob_score = self.loss_(y[~sample_mask],
y_pred[~sample_mask],
sample_weight[~sample_mask])
else:
subsample_weight = sample_weight
residuals = self.loss_.negative_gradient(
y, y_pred, sample_weight=sample_weight)
best_learner = _fit_stage_componentwise(Xi, residuals,
subsample_weight)
self.estimators_.append(best_learner)
if do_dropout:
drop_model, n_dropped = _sample_binomial_plus_one(
self.dropout_rate, num_iter + 1, random_state)
scale[num_iter] = 1. / (n_dropped + 1.)
y_pred[:] = 0
for m in range(num_iter + 1):
if drop_model[m] == 1:
scale[m] *= n_dropped / (n_dropped + 1.)
else:
y_pred += self.learning_rate * scale[
m] * self.estimators_[m].predict(Xi)
else:
y_pred += self.learning_rate * best_learner.predict(Xi)
# track deviance (= loss)
if do_oob:
self.train_score_[num_iter] = self.loss_(
y[sample_mask], y_pred[sample_mask],
sample_weight[sample_mask])
self.oob_improvement_[num_iter] = (
old_oob_score -
self.loss_(y[~sample_mask], y_pred[~sample_mask],
sample_weight[~sample_mask]))
else:
# no need to fancy index w/ no subsampling
self.train_score_[num_iter] = self.loss_(
y, y_pred, sample_weight)
if self.verbose > 0:
verbose_reporter.update(num_iter, self)