def main():
X, y = utils.load_10x(path('SC02'), 'SC02')
best_model = catboost.CatBoostClassifier(
l2_leaf_reg=3,
learning_rate=0.445,
depth=10,
iterations=200,
random_seed=42,
logging_level='Silent',
loss_function='MultiClass',
eval_metric='TotalF1',
thread_count=20,
)
model_path = os.path.join(CUR_DIR, 'sc02-best-model.cbm')
if os.path.exists(model_path):
best_model.load_model(model_path)
else:
best_model.fit(X, y, [X.shape[1] - 1])
best_model.save_model(model_path)
sc03, _ = utils.load_10x(path('SC03'), 'SC03')
sc03_preds = predict(best_model, X.columns, sc03)
sc03_preds.to_csv(os.path.join(CUR_DIR, 'sc03-preds.csv'))
importances = pd.DataFrame(best_model._feature_importance, X.columns)
importances[importances[0] > 0].sort_values(0, ascending=False).to_csv(
os.path.join(CUR_DIR, 'sc02-model-features.csv'))
def process():
X, _ = utils.load_10x(os.path.join(ROOT, 'SC01'), 'SC01v2')
X.drop(columns=['Batch'], inplace=True)
predictions = pd.read_csv(os.path.join(CUR_DIR, 'sc01-best-preds.csv'),
index_col=0)
predictions.index = predictions.index.str.replace('-1', '')
y = predictions.loc[X.index, :].max(axis=1)
params_space = {
'l2_leaf_reg': [1, 3, 5, 7, 9],
'learning_rate': np.linspace(1e-3, 8e-1, num=10),
'depth': [6, 8, 10],
}
score, best_params = find_best_params(X, y, params_space)
print('Best params score: {}'.format(score))
print('Best params: {}'.format(repr(best_params)))
model = catboost.CatBoostRegressor(
iterations=500,
learning_rate=best_params['learning_rate'],
depth=best_params['depth'],
l2_leaf_reg=best_params['l2_leaf_reg'],
random_seed=42,
logging_level='Silent',
thread_count=20,
loss_function='MAE',
)
model.fit(X, y)
model.save_model(os.path.join(CUR_DIR, 'sc01-pred-score.cbm'))
importances = pd.DataFrame(model._feature_importance, X.columns)
importances[importances[0] > 0].sort_values(0, ascending=False).to_csv(
os.path.join(CUR_DIR, 'sc01-pred-score-features.csv'))
def main():
X, _ = utils.load_10x(path('SC02'), 'SC02v2')
best_model = catboost.CatBoostClassifier(
l2_leaf_reg=7,
learning_rate=0.622,
depth=10,
iterations=200,
random_seed=42,
logging_level='Silent',
loss_function='MultiClass',
eval_metric='TotalF1',
thread_count=20,
)
model_path = os.path.join(CUR_DIR, 'sc02v2-best-model.cbm')
best_model.load_model(model_path)
sc01, _ = utils.load_10x(path('SC01'), 'SC01v2')
sc01_preds = predict(best_model, X.columns, sc01)
sc01_preds.to_csv(os.path.join(CUR_DIR, 'sc01-preds.csv'))
def main():
runs = []
for i in range(1, 6):
runs.append(
pd.read_csv(os.path.join(CUR_DIR,
'./search-{}-of-5.csv'.format(i)),
index_col=0))
runs = pd.concat(runs, ignore_index=True)
best_row = runs.score.idxmax(axis=0)
best_params = eval(runs.params[best_row])
X, y = utils.load_10x(path('SC03'), 'SC03')
best_model = catboost.CatBoostClassifier(
l2_leaf_reg=best_params['l2_leaf_reg'],
learning_rate=best_params['learning_rate'],
depth=best_params['depth'],
iterations=200,
random_seed=42,
logging_level='Silent',
loss_function='MultiClass',
eval_metric='TotalF1',
thread_count=20,
)
model_path = os.path.join(CUR_DIR, 'sc03-model.cbm')
if os.path.exists(model_path):
best_model.load_model(model_path)
else:
best_model.fit(X, y, [X.shape[1] - 1])
best_model.save_model(model_path)
importances = pd.DataFrame(best_model._feature_importance, X.columns)
importances[importances[0] > 0].sort_values(0, ascending=False).to_csv(
os.path.join(CUR_DIR, 'sc03-features.csv'))
sc01, _ = utils.load_10x(path('SC01'), 'SC01v2')
sc01_preds = predict(best_model, X.columns, sc01)
sc01_preds.to_csv(os.path.join(CUR_DIR, 'sc01-preds.csv'))
sc02, _ = utils.load_10x(path('SC02'), 'SC02v2')
sc02_preds = predict(best_model, X.columns, sc02)
sc02_preds.to_csv(os.path.join(CUR_DIR, 'sc02-preds.csv'))
def main(splits, current_split):
X, y = utils.load_10x(path('SC02'), 'SC02v2')
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.1, stratify=y, random_state=42,
)
params_space = {
'l2_leaf_reg': [1, 3, 5, 7, 9],
'learning_rate': np.linspace(1e-3, 8e-1, num=10),
'depth': [6, 8, 10],
}
score, best_params = catboost_GridSearchCV(X_train, y_train, params_space, cv=5,
splits=splits, current_split=current_split)
print(score)
print(best_params)
def main(splits, current_split):
X, y = utils.load_10x(path('SC03'), 'SC03')
params_space = {
'l2_leaf_reg': [1, 3, 5, 7, 9],
'learning_rate': np.linspace(1e-3, 8e-1, num=10),
'depth': [6, 8, 10],
}
record = []
score, best_params = catboost_GridSearchCV(X,
y,
params_space,
record,
cv=5,
splits=splits,
current_split=current_split)
pd.DataFrame(record, columns=['params', 'score', 'time']).to_csv(
os.path.join(CUR_DIR,
'search-{}-of-{}.csv'.format(current_split, splits)))
def get_query():
global _query
if '_query' not in globals():
_query = utils.load_10x(os.path.join(ROOT, 'SC03'), 'SC03')
return _query
def get_reference():
global _reference
if '_reference' not in globals():
_reference = utils.load_10x(os.path.join(ROOT, 'SC02'), 'SC02v2')
return _reference