def model_train_cv_parallel(model, X, y, cv=None, proba=True, n_jobs=8):
if cv is None:
cv = skl.cross_validation.KFold(len(y),
n_folds=16,
shuffle=True,
random_state=1234)
return pd.concat(
Parallel(n_jobs=n_jobs, verbose=True)(
delayed(model_train_cv)(fold, model, X, y, proba) for fold in cv))
def tune_gamma(self):
scores = self.score_gamma()
best_score = scores.iloc[0]
self.best_gamma = best_score['gamma']
gamma_range = self.detail_range(self.best_gamma, 0.05, min=0)
scores2 = self.score_gamma(gamma_range)
best_score2 = scores2.iloc[0]
self.best_gamma = best_score2['gamma']
scores = pd.concat([scores, scores2], ignore_index=True)
return scores.drop_duplicates().sort_values('score', ascending=self.metric_ascending)
def tune_gamma(self):
scores = self.score_gamma()
best_score = scores.iloc[0]
self.best_gamma = best_score['gamma']
gamma_range = self.detail_range(self.best_gamma, 0.05, min=0)
scores2 = self.score_gamma(gamma_range)
best_score2 = scores2.iloc[0]
self.best_gamma = best_score2['gamma']
scores = pd.concat([scores, scores2], ignore_index=True)
return scores.drop_duplicates().sort_values(
'score', ascending=self.metric_ascending)
def tune_sample(self):
scores = self.score_samples()
best_score = scores.iloc[0]
self.best_subsample = best_score['subsample']
self.best_colsample_bytree = best_score['colsample_bytree']
subsample_range = self.detail_range(self.best_subsample, 0.05, max=1)
colsample_bytree_range = self.detail_range(self.best_colsample_bytree, 0.05, max=1)
scores2 = self.score_samples(subsample_range, colsample_bytree_range)
best_score2 = scores2.iloc[0]
self.best_subsample = best_score2['subsample']
self.best_colsample_bytree = best_score2['colsample_bytree']
scores = pd.concat([scores, scores2], ignore_index=True)
return scores.drop_duplicates().sort_values('score', ascending=self.metric_ascending)
def tune_depth(self):
scores = self.score_depth()
best_score = scores.iloc[0]
self.best_max_depth = int(best_score['max_depth'])
self.best_min_child_weight = int(best_score['min_child_weight'])
max_depth_range = self.detail_range(self.best_max_depth, 1, min=1)
min_child_weight_range = self.detail_range(self.best_min_child_weight, 1, min=1)
scores2 = self.score_depth(max_depth_range, min_child_weight_range)
best_score2 = scores2.iloc[0]
self.best_max_depth = int(best_score2['max_depth'])
self.best_min_child_weight = int(best_score2['min_child_weight'])
scores = pd.concat([scores, scores2], ignore_index=True)
return scores.drop_duplicates().sort_values('score', ascending=self.metric_ascending)
def tune_sample(self):
scores = self.score_samples()
best_score = scores.iloc[0]
self.best_subsample = best_score['subsample']
self.best_colsample_bytree = best_score['colsample_bytree']
subsample_range = self.detail_range(self.best_subsample, 0.05, max=1)
colsample_bytree_range = self.detail_range(self.best_colsample_bytree,
0.05,
max=1)
scores2 = self.score_samples(subsample_range, colsample_bytree_range)
best_score2 = scores2.iloc[0]
self.best_subsample = best_score2['subsample']
self.best_colsample_bytree = best_score2['colsample_bytree']
scores = pd.concat([scores, scores2], ignore_index=True)
return scores.drop_duplicates().sort_values(
'score', ascending=self.metric_ascending)
def tune_depth(self):
scores = self.score_depth()
best_score = scores.iloc[0]
self.best_max_depth = int(best_score['max_depth'])
self.best_min_child_weight = int(best_score['min_child_weight'])
max_depth_range = self.detail_range(self.best_max_depth, 1, min=1)
min_child_weight_range = self.detail_range(self.best_min_child_weight,
1,
min=1)
scores2 = self.score_depth(max_depth_range, min_child_weight_range)
best_score2 = scores2.iloc[0]
self.best_max_depth = int(best_score2['max_depth'])
self.best_min_child_weight = int(best_score2['min_child_weight'])
scores = pd.concat([scores, scores2], ignore_index=True)
return scores.drop_duplicates().sort_values(
'score', ascending=self.metric_ascending)
def model_train_cv_parallel(model, X, y, cv=None, proba=True, n_jobs=8):
if cv is None:
cv = skl.cross_validation.KFold(len(y), n_folds=16, shuffle=True, random_state=1234)
return pd.concat(Parallel(n_jobs=n_jobs, verbose=True)(delayed(model_train_cv)(fold, model, X, y, proba) for fold in cv))
def synergy_calc_cv_parallel(heroes, y, cv):
return pd.concat(Parallel(n_jobs=8, verbose=True)(delayed(synergy_calc_cv)(fold, heroes, y) for fold in cv))