Used to generate trained GB models with different train-test splits.
"""
np.random.seed(100)
df = pickle.load(open('../../data/pairs_pdos.pkl'))
if True:
X,y = train_prep_pdos(df,include_WF=True,dE=0.1)
model_type = 'pdos'
else:
X,y = train_prep(df,include_WF=True)
model_type = 'moments'
if False:
X_train, X_dev, X_test, y_train, y_dev, y_test = split_by_cols(df,X,y,['comp','ads_a','ads_b'])
split_type = 'comp_rxn'
elif True:
X_train, X_dev, X_test, y_train, y_dev, y_test = split_by_cols(df,X,y,['comp'])
split_type = 'comp'
elif False:
X_train, X_dev, X_test, y_train, y_dev, y_test = split_by_cols(df,X,y,['ads_a','ads_b'])
split_type = 'rxn'
else:
X_train, X_test, y_train, y_test = train_test_split(X,y,test_size=0.3)
X_dev, X_test, y_dev, y_test = train_test_split(X_test,y_test,test_size=0.5)
split_type = 'random'
model = CatBoostRegressor(loss_function='MAE',iterations=1.5e4)
model.fit(X_train,y_train,eval_set=(X_dev,y_dev))
model.save_model('%s_%s.cbm'%(model_type,split_type))
return mae
if __name__ == '__main__':
df = pickle.load(open('data/pairs_pdos.pkl'))
features = 'moments' # #pdos,'moments'
bayes = True
#Feature Selection
if features == 'moments':
X, y = train_prep(df)
elif features == 'pdos':
X, y = train_prep_pdos(df, stack=False, include_WF=False, dE=0.1)
X_train, X_dev, X_test, y_train, y_dev, y_test, groups = split_by_cols(
df, X, y, ['comp', 'ads_a', 'ads_b'], ret_groups=True)
rf = ensemble.RandomForestRegressor(n_estimators=100)
group_kfold = GroupKFold(n_splits=3)
#print(X_train.shape[1]),np.sqrt(X_train.shape[1])
if bayes:
random_grid = { #'n_estimators': (5,100),
'max_features': (int(np.sqrt(X_train.shape[1])), X_train.shape[1]),
'max_depth': (5, 50),
'min_samples_split': (2, 10),
'min_samples_leaf': (2, 5),
'bootstrap': [True, False]
}
df = pickle.load(open('../../data/pairs_pdos.pkl'))
X, y = train_prep_pdos(df, include_WF=False, stack=True)
for split_type in ['comp', 'rxn', 'comp_rxn', 'random']:
print split_type
if split_type == 'comp':
cols = ['comp']
elif split_type == 'rxn':
cols = ['ads_a', 'ads_b']
elif split_type == 'comp_rxn':
cols = ['comp', 'ads_a', 'ads_b']
else:
cols = None
if cols != None:
X_train, X_dev, X_test, y_train, y_dev, y_test = split_by_cols(
df, X, y, cols)
else:
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.3)
X_dev, X_test, y_dev, y_test = train_test_split(X_test,
y_test,
test_size=0.5)
np.save('X_%s_train.npy' % split_type, X_train)
np.save('X_%s_dev.npy' % split_type, X_dev)
np.save('X_%s_test.npy' % split_type, X_test)
np.save('y_%s_train.npy' % split_type, y_train)
np.save('y_%s_dev.npy' % split_type, y_dev)
np.save('y_%s_test.npy' % split_type, y_test)