def cross_validate(self, x_train, y_train, x_test, folds, verbose_eval=False):
if int(verbose_eval) > 0:
print("Validation starts")
print(x_train.shape)
for i, (train_idx, test_idx) in enumerate(folds):
print("Fold %i" % (i + 1))
x_trn = x_train.iloc[train_idx,]
y_trn = y_train[train_idx]
x_holdout = x_train.iloc[test_idx,]
y_holdout = y_train[test_idx]
eval_set = [(x_holdout, y_holdout)]
# match to a proper sklearn .fit function
self.fit_func(x_trn, y_trn)
self.y_trn[test_idx] = self.clf.predict_proba(x_holdout)[:, 1]
self.holdout_gini = eval_gini_normalized(y_holdout, self.y_trn[test_idx])
self.y_tst_tmp[:, i] = self.clf.predict_proba(x_test)[:, 1]
gc.collect()
self.trn_gini = eval_gini_normalized(y_train, self.y_trn)
self.y_tst = np.mean(self.y_tst_tmp, axis=1)
if int(verbose_eval) > 0:
print ("CV score for train cv set: %f" % self.trn_gini)
def cross_validate_xgb(self,
x_train,
y_train,
x_test,
folds,
verbose_eval=False):
''' cross validation for xgb not with sklearn api
'''
if int(verbose_eval) > 0:
print("Validation starts")
print(x_train.shape)
for i, (train_idx, test_idx) in enumerate(folds):
if int(verbose_eval) > 0:
print("Fold %i" % (i + 1))
x_trn = x_train.iloc[train_idx, ]
y_trn = y_train[train_idx]
x_holdout = x_train.iloc[test_idx, ]
y_holdout = y_train[test_idx]
if self.upsam:
x_trn, y_trn = self.upsampling(x_trn, y_trn)
d_trn = xgb.DMatrix(x_trn, label=y_trn)
d_tst = xgb.DMatrix(x_test)
d_holdout = xgb.DMatrix(x_holdout, label=y_holdout)
watchlist = [(d_holdout, 'holdout')]
del x_trn, y_trn, x_holdout
gc.collect()
self.clf = xgb.train(params=self.params,
dtrain=d_trn,
num_boost_round=self.max_round,
evals=watchlist,
feval=gini_xgb,
maximize=True,
verbose_eval=verbose_eval,
early_stopping_rounds=50)
self.y_trn[test_idx] = self.clf.predict(d_holdout)
self.holdout_gini = eval_gini_normalized(y_holdout,
self.y_trn[test_idx])
self.y_tst_tmp[:, i] = self.clf.predict(d_tst)
self.fscore = pd.concat(
[self.fscore,
pd.Series(self.clf.get_fscore(), name=i)],
axis=1)
del d_trn, d_holdout, y_holdout
gc.collect()
self.trn_gini = eval_gini_normalized(y_train, self.y_trn)
self.y_tst = np.mean(self.y_tst_tmp, axis=1)
if int(verbose_eval) > 0:
print("CV score for train cv set: %f" % self.trn_gini)
def bagging(self, x_train, y_train, x_test, idx, verbose_eval=False):
random_seeds = [177, 47, 8243, 5210, 1]
print("Bagging of %i" % self.bags)
trn_series = np.zeros((self.len_trn, self.bags))
tst_series = np.zeros((self.len_tst, self.bags))
for i in range(self.bags):
print("Bag %i" % (i + 1))
folds = list(KFold(
n_splits=self.n_splits,
shuffle=True,
random_state=random_seeds[i]
).split(x_train, y_train))
self.methods[idx](x_train, y_train, x_test, folds, verbose_eval)
trn_series[:, i] = self.y_trn
tst_series[:, i] = self.y_tst
self.y_tst_tmp2[:, (i*5):((i+1)*5)] = self.y_tst_tmp
self.trn_gini_bags[i] = self.trn_gini
# Reset all values (not necessary but)
self.y_trn = np.zeros(self.len_trn)
self.y_tst = np.zeros(self.len_tst)
self.y_tst_tmp = np.zeros((self.len_tst, self.n_splits))
self.y_trn = np.mean(trn_series, axis=1)
self.y_tst = np.mean(tst_series, axis=1)
self.trn_gini = eval_gini_normalized(y_train, self.y_trn)
print ("CV score among different bags:\n")
print ("CV score among different bags:\n")
print self.trn_gini_bags
print ("CV score var"),
print np.var(self.trn_gini_bags)
print ("\nCV score for %i aver cv set: %f" % (self.bags, self.trn_gini))
def log_reg(trn_series, tst_series, y_train):
clf = LogisticRegression()
clf.fit(trn_series, y_train)
score = eval_gini_normalized(y_train, clf.predict_proba(trn_series)[:, 1])
y_tst = clf.predict_proba(tst_series)[:, 1]
return score, y_tst
def mean(trn_series, tst_series, y_train):
score = eval_gini_normalized(y_train, np.mean(trn_series, axis=1))
y_tst = np.mean(tst_series, axis=1)
return score, y_tst