def build_target(self):
y, ff = build_target_safe(self.target, self.data)
self.y = y
self.negatives = y[~y.astype('bool')].index
self.positives = y[y.astype('bool')].index
self.n_positives = len(self.positives)
self.n_negatives = len(self.negatives)
def cross_validate(model_def, data, folds, repeat=1):
"""
"""
results = []
if isinstance(folds, int):
folds = make_default_folds(num_folds=folds, data=data)
for i in range(repeat):
for fold in folds:
if len(fold) == 2:
train_index, test_index = fold
prep_index = None
elif len(fold) == 3:
train_index, test_index, prep_index = fold
else:
raise ValueError("Fold is not of right dimension (%d, not 2 or 3)"%len(fold))
assert len(train_index & test_index) == 0, "train and test overlap!!! %s, %s" % (train_index, test_index)
x_train, y_train, fitted_model = fit_model(model_def, data, prep_index, train_index)
test_data = data.loc[test_index]
x_test, y_test = generate_test(model_def, test_data, fitted_model)
assert len(x_train.index & x_test.index) == 0, "train and test overlap!!! %s" % (x_train.index & x_test.index)
y_preds = predict(fitted_model, x_test)
if model_def.evaluation_target is not None:
y_test, ff = build_target_safe(model_def.evaluation_target, test_data)
result = Result(x_train, x_test, y_train, y_test, y_preds, model_def, fitted_model, data)
results.append(result)
# for reporter in reporters:
# reporter.update(result)
return results
def generate_train(model_def, data, prep_index=None, train_index=None):
# create training set
data, prep_index, train_index = filter_data_and_indexes(model_def, data, prep_index, train_index)
x_train, fitted_features = build_featureset_safe(model_def.features, data, prep_index, train_index)
y_train, fitted_target = build_target_safe(model_def.target, data, prep_index, train_index)
x_train = x_train.reindex(train_index)
y_train = y_train.reindex(train_index)
return x_train, y_train, fitted_features, fitted_target
def predict(fitted_model, x_data):
model_def = fitted_model.model_def
predictions = fitted_model.fitted_estimator.predict(x_data)
predictions = pd.Series(predictions, index=x_data.index)
if model_def.evaluation_transformation is not None:
x_data[model_def.predictions_name] = predictions
predictions, ff = build_target_safe(model_def.evaluation_transformation, x_data)
del x_data[model_def.predictions_name]
return predictions
def _train(self, train_datas):
train_data = concat(train_datas, axis=1)
y, ff = build_target_safe(self.target, self.data)
y = reindex_safe(y, train_data.index)
arg = self.threshold_arg
if arg is None:
arg = self.n_keep
cols = self.selector.select(train_data, y, arg)
return cols
def _train(self, train_datas):
train_data = concat(train_datas, axis=1)
y, ff = build_target_safe(self.target, self.data)
y = reindex_safe(y, train_data.index)
arg = self.threshold_arg
if arg is None:
arg = self.n_keep
cols = self.selector.select(train_data, y, arg)
return cols
def predict(fitted_model, x_data):
model_def = fitted_model.model_def
predictions = fitted_model.fitted_estimator.predict(x_data)
predictions = pd.Series(predictions, index=x_data.index)
if model_def.evaluation_transformation is not None:
x_data[model_def.predictions_name] = predictions
predictions, ff = build_target_safe(
model_def.evaluation_transformation, x_data)
del x_data[model_def.predictions_name]
return predictions
def generate_train(model_def, data, prep_index=None, train_index=None):
# create training set
data, prep_index, train_index = filter_data_and_indexes(
model_def, data, prep_index, train_index)
x_train, fitted_features = build_featureset_safe(model_def.features, data,
prep_index, train_index)
y_train, fitted_target = build_target_safe(model_def.target, data,
prep_index, train_index)
x_train = x_train.reindex(train_index)
y_train = y_train.reindex(train_index)
return x_train, y_train, fitted_features, fitted_target
def _train(self, train_data):
y, ff = build_target_safe(self.target, train_data)
vc = train_data[self.group_by].value_counts()
keys = [k for k, v in vc.iterkv() if v >= self.min_sample]
train_data['__grouping'] = train_data[self.group_by].map(lambda x: x if x in keys else '__other')
train_data['__target'] = y
vals = train_data.groupby('__grouping').agg({'__target': self.func})['__target'].to_dict()
logging.debug("Preparing Target Aggregations:")
logging.debug(str(vals.items()[:10]))
del train_data['__target']
del train_data['__grouping']
return (keys, vals)
def fit_model(model_def, data, prep_index=None, train_index=None):
# create training set
x_train, fitted_features = build_featureset_safe(model_def.features, data, prep_index, train_index)
y_train, fitted_target = build_target_safe(model_def.target, data, prep_index, train_index)
# fit estimator
model_def.estimator.fit(x_train, y_train)
# unnecesary?
fitted_estimator = FittedEstimator(model_def.estimator, x_train, y_train)
fitted_model = FittedModel(model_def, fitted_features, fitted_target, fitted_estimator)
return x_train, y_train, fitted_model
def _train(self, train_data):
y, ff = build_target_safe(self.target, train_data)
train_data['__target'] = y
global_value = self.func(y)
if self.regularize:
keys = train_data[self.group_by].unique()
f = lambda x: (self.func(x) * x.size + global_value * self.min_sample) / (x.size + self.min_sample)
vals = train_data.groupby(self.group_by).agg({'__target': f})['__target'].to_dict()
else:
vc = train_data[self.group_by].value_counts()
keys = [k for k, v in vc.iterkv() if v >= self.min_sample]
train_data['__grouping'] = train_data[self.group_by].map(lambda x: x if x in keys else '__other')
vals = train_data.groupby('__grouping').agg({'__target': self.func})['__target'].to_dict()
del train_data['__grouping']
if '__other' not in vals:
vals['__other'] = global_value
logging.debug("Preparing Target Aggregations:")
logging.debug(str(vals.items()[:10]))
del train_data['__target']
return vals
def cross_validate(model_def, data, folds, repeat=1):
"""
"""
results = []
if isinstance(folds, int):
folds = make_default_folds(num_folds=folds, data=data)
for i in range(repeat):
for fold in folds:
if len(fold) == 2:
train_index, test_index = fold
prep_index = None
elif len(fold) == 3:
train_index, test_index, prep_index = fold
else:
raise ValueError(
"Fold is not of right dimension (%d, not 2 or 3)" %
len(fold))
assert len(
train_index
& test_index) == 0, "train and test overlap!!! %s, %s" % (
train_index, test_index)
x_train, y_train, fitted_model = fit_model(model_def, data,
prep_index, train_index)
test_data = data.loc[test_index]
x_test, y_test = generate_test(model_def, test_data, fitted_model)
assert len(x_train.index
& x_test.index) == 0, "train and test overlap!!! %s" % (
x_train.index & x_test.index)
y_preds = predict(fitted_model, x_test)
if model_def.evaluation_target is not None:
y_test, ff = build_target_safe(model_def.evaluation_target,
test_data)
result = Result(x_train, x_test, y_train, y_test, y_preds,
model_def, fitted_model, data)
results.append(result)
# for reporter in reporters:
# reporter.update(result)
return results
def _train(self, train_data):
y, ff = build_target_safe(self.target, train_data)
train_data['__target'] = y
global_value = self.func(y)
if self.regularize:
keys = train_data[self.group_by].unique()
f = lambda x: (self.func(x) * x.size + global_value * self.
min_sample) / (x.size + self.min_sample)
vals = train_data.groupby(self.group_by).agg(
{'__target': f})['__target'].to_dict()
else:
vc = train_data[self.group_by].value_counts()
keys = [k for k, v in vc.iterkv() if v >= self.min_sample]
train_data['__grouping'] = train_data[self.group_by].map(
lambda x: x if x in keys else '__other')
vals = train_data.groupby('__grouping').agg(
{'__target': self.func})['__target'].to_dict()
del train_data['__grouping']
if '__other' not in vals:
vals['__other'] = global_value
logging.debug("Preparing Target Aggregations:")
logging.debug(str(vals.items()[:10]))
del train_data['__target']
return vals
def generate_train(model_def, data, prep_index=None, train_index=None):
# create training set
x_train, fitted_features = build_featureset_safe(model_def.features, data, prep_index, train_index)
y_train, fitted_target = build_target_safe(model_def.target, data, prep_index, train_index)
return x_train, y_train, fitted_features, fitted_target
