def run_predict_cv(self) -> None:
"""クロスバリデーションで学習した各foldのモデルの平均により、テストデータの予測を行う
あらかじめrun_train_cvを実行しておく必要がある
"""
logger.info(f'{self.run_name} - start prediction cv')
X_test = self.X_test
preds = []
show_feature_importance = 'LGBM' in str(self.model_cls)
if show_feature_importance:
feature_importances = pd.DataFrame()
# 各foldのモデルで予測を行う
for i_fold in range(self.cv.n_splits):
logger.info(f'{self.run_name} - start prediction fold:{i_fold}')
model = self.build_model(i_fold)
model.load_model()
pred = model.predict(X_test)
preds.append(pred)
logger.info(f'{self.run_name} - end prediction fold:{i_fold}')
if show_feature_importance:
feature_importances = pd.concat(
[feature_importances,
model.feature_importance(X_test)],
axis=0)
# 予測の平均値を出力する
pred_avg = np.mean(preds, axis=0)
# 予測結果の保存
Data.dump(pred_avg, f'../output/pred/{self.run_name}-test.pkl')
logger.info(f'{self.run_name} - end prediction cv')
# 特徴量の重要度
if show_feature_importance:
aggs = feature_importances.groupby('Feature').mean().sort_values(
by="importance", ascending=False)
cols = aggs[:200].index
pd.DataFrame(aggs.index).to_csv(
f'../output/importance/{self.run_name}-fi.csv', index=False)
best_features = feature_importances.loc[
feature_importances.Feature.isin(cols)]
plt.figure(figsize=(14, 26))
sns.barplot(x="importance",
y="Feature",
data=best_features.sort_values(by="importance",
ascending=False))
plt.title('LightGBM Features (averaged over folds)')
plt.tight_layout()
plt.savefig(f'../output/importance/{self.run_name}-fi.png')
plt.show()
# mlflow
mlflow.start_run(run_id=self.run_id)
log_artifact(f'../output/importance/{self.run_name}-fi.png')
mlflow.end_run()
def make_predictions(data: list, weights: list):
pred = 0
for i, d in enumerate(data):
if i < len(weights):
pred += d[1] * weights[i]
else:
pred += d[1] * (1 - sum(weights))
Data.dump(pred, f'../output/pred/{run_name}-test.pkl')
return pred
def f(x):
pred = 0
for i, d in enumerate(data):
if i < len(x):
pred += d[0] * x[i]
else:
pred += d[0] * (1 - sum(x))
score = -1 * roc_auc_score(y_train, pred)
Data.dump(pred, f'../output/pred/{run_name}-train.pkl')
return score
def f(x):
pred = 0
for i, d in enumerate(data):
if i < len(x):
pred += d[0] * x[i]
else:
pred += d[0] * (1 - sum(x))
score = np.sqrt(mean_squared_error(y_train, pred))
Data.dump(pred, f'../output/pred/{run_name}-train.pkl')
return score
def run_train_cv(self) -> None:
"""クロスバリデーションでの学習・評価を行う
学習・評価とともに、各foldのモデルの保存、スコアのログ出力についても行う
"""
logger.info(f'{self.run_name} - start training cv')
scores = []
va_idxes = []
preds = []
# 各foldで学習を行う
for i_fold in range(self.n_fold):
# 学習を行う
logger.info(f'{self.run_name} fold {i_fold} - start training')
model, va_idx, va_pred, score = self.train_fold(i_fold)
logger.info(
f'{self.run_name} fold {i_fold} - end training - score {score}'
)
# モデルを保存する
model.save_model()
# 結果を保持する
va_idxes.append(va_idx)
scores.append(score)
preds.append(va_pred)
# 各foldの結果をまとめる
va_idxes = np.concatenate(va_idxes)
order = np.argsort(va_idxes)
preds = np.concatenate(preds, axis=0)
preds = preds[order]
logger.info(
f'{self.run_name} - end training cv - score {np.mean(scores)}')
# 予測結果の保存
Data.dump(preds, f'../output/pred/{self.run_name}-train.pkl')
# 評価結果の保存
logger.result_scores(self.run_name, scores)
def save_as_pickle(
train: pd.DataFrame,
test: pd.DataFrame,
target_col: str,
exp_id: str,
output_dir: str = '../input') -> Tuple[pd.DataFrame, pd.DataFrame]:
"""Save X_train, X_test and y_train as pickel format
Args:
train (pd.DataFrame): train
test (pd.DataFrame): test
target_col (str): target column
exp_id (str): experiment id
output_dir (str, optional): output directory. Defaults to '../input'.
Returns:
Tuple[pd.DataFrame, pd.DataFrame]: train, test
"""
X_train = train.drop(target_col, axis=1)
y_train = train[target_col]
if target_col in test.columns:
X_test = test.drop(target_col, axis=1)
else:
X_test = test
Data.dump(X_train, join(output_dir, f"X_train_{exp_id}.pkl"))
Data.dump(y_train, join(output_dir, f"y_train_{exp_id}.pkl"))
Data.dump(X_test, join(output_dir, f"X_test_{exp_id}.pkl"))
def save_as_pickle(train: pd.DataFrame, test: pd.DataFrame,
col_definition: dict, option: dict):
"""
col_definition: target_col
option: exp_id
"""
X_train = train.drop(col_definition['target_col'], axis=1)
y_train = train[col_definition['target_col']]
if col_definition['target_col'] in test.columns:
X_test = test.drop(col_definition['target_col'], axis=1)
else:
X_test = test
Data.dump(X_train, join('../input', f"X_train{option['exp_id']}.pkl"))
Data.dump(y_train, join('../input', 'y_train.pkl'))
Data.dump(X_test, join('../input', f"X_test{option['exp_id']}.pkl"))
def save_as_pickle(train: pd.DataFrame, test: pd.DataFrame,
col_definition: dict, option: dict):
"""
col_definition: target_col
option: output_dir, exp_id
"""
X_train = train.drop(col_definition['target_col'], axis=1)
y_train = train[col_definition['target_col']]
if col_definition['target_col'] in test.columns:
X_test = test.drop(col_definition['target_col'], axis=1)
else:
X_test = test
Data.dump(X_train,
join(option['output_dir'], f"X_train_{option['exp_id']}.pkl"))
Data.dump(y_train,
join(option['output_dir'], f"y_train_{option['exp_id']}.pkl"))
Data.dump(X_test,
join(option['output_dir'], f"X_test_{option['exp_id']}.pkl"))
def save_model(self) -> None:
model_path = os.path.join("../output/model",
f"{self.run_fold_name}.model")
os.makedirs(os.path.dirname(model_path), exist_ok=True)
# best_ntree_limitが消えるのを防ぐため、pickleで保存することとした
Data.dump(self.model, model_path)
'agg': ['mean', 'sum', 'median', 'min', 'max', 'var', 'std']
},
],
nunique_dict=[
{
'key': ['Sex'],
'var': ['SibSp'],
'agg': ['nunique']
},
{
'key': ['Sex'],
'var': ['Cabin'],
'agg': ['nunique']
},
])
print(X_train.shape, X_test.shape)
unique_cols, duplicated_cols, high_corr_cols = detect_delete_cols(
X_train, X_test, escape_col=categorical_cols, threshold=0.99)
X_train.drop(unique_cols + duplicated_cols + high_corr_cols,
axis=1,
inplace=True)
X_test.drop(unique_cols + duplicated_cols + high_corr_cols,
axis=1,
inplace=True)
print(X_train.shape, X_test.shape)
Data.dump(X_train, output_dir + 'X_train_fe000.pkl')
Data.dump(X_test, output_dir + 'X_test_fe000.pkl')
Data.dump(y_train, output_dir + 'y_train_fe000.pkl')
import pandas as pd
from ayniy.utils import Data
INPUT_DIR = "../input/data_v2/"
if __name__ == "__main__":
train = pd.read_csv(INPUT_DIR + "train.csv")
test = pd.read_csv(INPUT_DIR + "test.csv")
Data.dump(train.drop("score", axis=1), "../input/X_train_fe000.pkl")
Data.dump(train["score"], "../input/y_train_fe000.pkl")
Data.dump(test, "../input/X_test_fe000.pkl")
def save_model(self) -> None:
model_path = os.path.join("../output/model", f"{self.run_fold_name}.model")
os.makedirs(os.path.dirname(model_path), exist_ok=True)
Data.dump(self.model, model_path)
def save_model(self):
model_path = os.path.join('../output/model', f'{self.run_fold_name}.model')
os.makedirs(os.path.dirname(model_path), exist_ok=True)
Data.dump(self.model, model_path)
fe_id = "fe000"
fe_name = f"{fe_id}_nn_small"
X_train = Data.load(f"../input/pickle/X_train_{fe_id}.pkl")
y_train = Data.load(f"../input/pickle/y_train_{fe_id}.pkl")
X_test = Data.load(f"../input/pickle/X_test_{fe_id}.pkl")
del_col = []
for c in X_train.columns:
X_train[c].fillna(-1, inplace=True)
X_test[c].fillna(-1, inplace=True)
try:
prep = StandardScaler()
X_train[c] = prep.fit_transform(X_train[[c]])
X_test[c] = prep.transform(X_test[[c]])
except ValueError:
del_col.append(c)
print(del_col)
print(len(del_col))
X_train.drop(del_col, axis=1, inplace=True)
X_test.drop(del_col, axis=1, inplace=True)
print(X_train.shape)
X_train = X_train.loc[:100]
y_train = y_train.loc[:100]
Data.dump(X_train, f"../input/pickle/X_train_{fe_name}.pkl")
Data.dump(y_train, f"../input/pickle/y_train_{fe_name}.pkl")
Data.dump(X_test, f"../input/pickle/X_test_{fe_name}.pkl")
fe005_tr = Data.load('../input/X_train_fe005.pkl')
fe005_te = Data.load('../input/X_test_fe005.pkl')
# fe001_tr = Data.load('../input/X_train_fe001.pkl')
# fe001_te = Data.load('../input/X_test_fe001.pkl')
# top10_tr = Data.load('../input/X_train_fe004_top10.pkl')
# top10_te = Data.load('../input/X_test_fe004_top10.pkl')
# top10_tr, top10_te = standerize(top10_tr, top10_te, {'encode_col': top10_tr.columns})
# print(top10_tr.head())
train_fitting_ef_add = pd.concat([fe005_tr, add_tr], axis=1)
test_fitting_ef_add = pd.concat([fe005_te, add_te], axis=1)
fe_name = 'fe005_add'
Data.dump(train_fitting_ef_add, f'../input/X_train_{fe_name}.pkl')
Data.dump(test_fitting_ef_add, f'../input/X_test_{fe_name}.pkl')
# train_fitting_ef.to_csv('../input/train_fitting_ef.csv', index=False)
# test_fitting_ef.to_csv('../input/test_fitting_ef.csv', index=False)
# spec_train = []
# for i, filename in enumerate(train['spectrum_filename'].values):
# spec_df = pd.read_csv(f'../input/spectrum_raw/{filename}', sep='\t', header=None)
# spec_train.append(spec_df[1])
# spec_tr = pd.concat(spec_train, axis=1)
# spec_tr = spec_tr.T
# spec_tr.columns = [f'spec_{i}' for i in range(spec_tr.shape[1])]
# spec_tr = spec_tr.reset_index(drop=True)
if __name__ == "__main__":
train = pd.read_csv(INPUT_DIR + "train.csv")
train = split_user_id(train)
train = add_user_ages(train)
train = add_user_purposes(train)
train = add_user_vecs(train)
train = add_user_strengths(train)
train = add_user_works(train)
train = add_user_skills(train)
train = add_user_educations(train)
pd.Series(train.columns).to_csv("../input/col_names.csv", index=False)
Data.dump(
train.drop(DELETE_COLS + ["score"], axis=1),
"../input/X_train_fe002.pkl",
)
Data.dump(train["score"], "../input/y_train_fe002.pkl")
del train
gc.collect()
test = pd.read_csv(INPUT_DIR + "test.csv")
test = split_user_id(test)
test = add_user_ages(test)
test = add_user_purposes(test)
test = add_user_vecs(test)
test = add_user_strengths(test)
test = add_user_works(test)
test = add_user_skills(test)
test = add_user_educations(test)
def run_train_cv(self) -> None:
"""クロスバリデーションでの学習・評価を行う
学習・評価とともに、各foldのモデルの保存、スコアのログ出力についても行う
"""
# mlflow
mlflow.set_experiment(self.exp_name)
mlflow.start_run(run_name=self.run_name)
logger.info(f"{self.run_name} - start training cv")
scores = []
va_idxes = []
preds = []
# Adversarial validation
if self.advanced and "adversarial_validation" in self.advanced:
X_train = self.X_train
X_test = self.X_test
X_train["target"] = 0
X_test["target"] = 1
X_train = pd.concat([X_train, X_test],
sort=False).reset_index(drop=True)
y_train = X_train["target"]
X_train.drop("target", axis=1, inplace=True)
X_test.drop("target", axis=1, inplace=True)
self.X_train = X_train
self.y_train = y_train
# 各foldで学習を行う
for i_fold in range(self.cv.n_splits):
# 学習を行う
logger.info(f"{self.run_name} fold {i_fold} - start training")
model, va_idx, va_pred, score = self.train_fold(i_fold)
logger.info(
f"{self.run_name} fold {i_fold} - end training - score {score}"
)
# モデルを保存する
model.save_model()
# 結果を保持する
va_idxes.append(va_idx)
scores.append(score)
preds.append(va_pred)
# 各foldの結果をまとめる
va_idxes = np.concatenate(va_idxes)
order = np.argsort(va_idxes)
preds = np.concatenate(preds, axis=0)
preds = preds[order]
if self.evaluation_metric == "log_loss":
cv_score = log_loss(self.y_train, preds, eps=1e-15, normalize=True)
elif self.evaluation_metric == "mean_absolute_error":
cv_score = mean_absolute_error(self.y_train, preds)
elif self.evaluation_metric == "rmse":
cv_score = np.sqrt(mean_squared_error(self.y_train, preds))
elif self.evaluation_metric == "auc":
cv_score = roc_auc_score(self.y_train, preds)
elif self.evaluation_metric == "prauc":
cv_score = average_precision_score(self.y_train, preds)
logger.info(f"{self.run_name} - end training cv - score {cv_score}")
# 予測結果の保存
Data.dump(preds, f"../output/pred/{self.run_name}-train.pkl")
# mlflow
self.run_id = mlflow.active_run().info.run_id
log_param("model_name", self.model_cls.__class__.__name__)
log_param("fe_name", self.fe_name)
log_param("train_params", self.params)
log_param("cv_strategy", str(self.cv))
log_param("evaluation_metric", self.evaluation_metric)
log_metric("cv_score", cv_score)
log_param(
"fold_scores",
dict(
zip([f"fold_{i}" for i in range(len(scores))],
[round(s, 4) for s in scores])),
)
log_param("cols_definition", self.cols_definition)
log_param("description", self.description)
mlflow.end_run()
import pandas as pd
from ayniy.utils import Data
if __name__ == '__main__':
ef_tr = pd.read_csv('../input/efficient_tr.csv')
ef_te = pd.read_csv('../input/efficient_te.csv')
fe001_top500_tr = Data.load('../input/X_train_fe001_top500.pkl')
fe001_top500_te = Data.load('../input/X_test_fe001_top500.pkl')
train_tag = pd.concat([fe001_top500_tr, ef_tr], axis=1)
test_tag = pd.concat([fe001_top500_te, ef_te], axis=1)
fe_name = 'fe001_top500_ef'
Data.dump(train_tag, f'../input/X_train_{fe_name}.pkl')
Data.dump(test_tag, f'../input/X_test_{fe_name}.pkl')
lbl = OrdinalEncoder(dtype='int')
train[col] = lbl.fit_transform(train[col].astype('str').fillna('-1').values.reshape(-1, 1))
test[col] = lbl.transform(test[col].astype('str').fillna('-1').values.reshape(-1, 1))
temp = pd.concat([train[[col]], test[[col]]], axis=0)
temp_mapping = temp.groupby(col).size() / len(temp)
temp['enc'] = temp[col].map(temp_mapping)
temp['enc'] = stats.rankdata(temp['enc'])
temp = temp.reset_index(drop=True)
train[f'rank_frqenc_{col}'] = temp[['enc']].values[:train.shape[0]]
test[f'rank_frqenc_{col}'] = temp[['enc']].values[train.shape[0]:]
test[col] = test[col].astype('category')
train[col] = train[col].astype('category')
drop_cols = list(set(drop_cols))
print(len(drop_cols))
train = train.drop(drop_cols, axis=1)
test = test.drop(drop_cols, axis=1)
train = reduce_mem_usage(train)
test = reduce_mem_usage(test)
gc.collect()
print(train.shape, test.shape)
test['encounter_id'] = test_id
test = test.sort_values('encounter_id').reset_index(drop=True)
fe_name = 'fe_siavrez'
Data.dump(train, f'../input/pickle/X_train_{fe_name}.pkl')
# Data.dump(y, f'../input/pickle/y_train_{fe_name}.pkl')
Data.dump(test.drop('encounter_id', axis=1), f'../input/pickle/X_test_{fe_name}.pkl')
def run_train_cv(self) -> None:
"""クロスバリデーションでの学習・評価を行う
学習・評価とともに、各foldのモデルの保存、スコアのログ出力についても行う
"""
# mlflow
mlflow.set_experiment(self.exp_name)
mlflow.start_run(run_name=self.run_name)
logger.info(f'{self.run_name} - start training cv')
scores = []
va_idxes = []
preds = []
# Adversarial validation
if self.advanced and 'adversarial_validation' in self.advanced:
X_train = self.X_train
X_test = self.X_test
X_train['target'] = 0
X_test['target'] = 1
X_train = pd.concat([X_train, X_test],
sort=False).reset_index(drop=True)
y_train = X_train['target']
X_train.drop('target', axis=1, inplace=True)
X_test.drop('target', axis=1, inplace=True)
self.X_train = X_train
self.y_train = y_train
# 各foldで学習を行う
for i_fold in range(self.cv.n_splits):
# 学習を行う
logger.info(f'{self.run_name} fold {i_fold} - start training')
model, va_idx, va_pred, score = self.train_fold(i_fold)
logger.info(
f'{self.run_name} fold {i_fold} - end training - score {score}'
)
# モデルを保存する
model.save_model()
# 結果を保持する
va_idxes.append(va_idx)
scores.append(score)
preds.append(va_pred)
# 各foldの結果をまとめる
va_idxes = np.concatenate(va_idxes)
order = np.argsort(va_idxes)
preds = np.concatenate(preds, axis=0)
preds = preds[order]
if self.evaluation_metric == 'log_loss':
cv_score = log_loss(self.y_train, preds, eps=1e-15, normalize=True)
elif self.evaluation_metric == 'mean_absolute_error':
cv_score = mean_absolute_error(self.y_train, preds)
elif self.evaluation_metric == 'rmse':
cv_score = np.sqrt(mean_squared_error(self.y_train, preds))
elif self.evaluation_metric == 'auc':
cv_score = roc_auc_score(self.y_train, preds)
elif self.evaluation_metric == 'prauc':
cv_score = average_precision_score(self.y_train, preds)
logger.info(f'{self.run_name} - end training cv - score {cv_score}')
# 予測結果の保存
Data.dump(preds, f'../output/pred/{self.run_name}-train.pkl')
# mlflow
self.run_id = mlflow.active_run().info.run_id
log_param('model_name', str(self.model_cls).split('.')[-1][:-2])
log_param('fe_name', self.fe_name)
log_param('train_params', self.params)
log_param('cv_strategy', str(self.cv))
log_param('evaluation_metric', self.evaluation_metric)
log_metric('cv_score', cv_score)
log_param(
'fold_scores',
dict(
zip([f'fold_{i}' for i in range(len(scores))],
[round(s, 4) for s in scores])))
log_param('cols_definition', self.cols_definition)
log_param('description', self.description)
mlflow.end_run()
ython select_features.py --n 100
"""
import argparse
import pandas as pd
from ayniy.utils import Data
parser = argparse.ArgumentParser()
parser.add_argument('--n')
args = parser.parse_args()
fe_id = 'fe005'
run_id = 'run046'
N_FEATURES = int(args.n)
fe_name = f'fe005_top{N_FEATURES}'
X_train = Data.load(f'../input/X_train_{fe_id}.pkl')
y_train = Data.load(f'../input/y_train_{fe_id}.pkl')
X_test = Data.load(f'../input/X_test_{fe_id}.pkl')
fi = pd.read_csv(
f'../output/importance/{run_id}-fi.csv')['Feature'][:N_FEATURES]
X_train = X_train[fi]
X_test = X_test[fi]
Data.dump(X_train, f'../input/X_train_{fe_name}.pkl')
Data.dump(y_train, f'../input/y_train_{fe_name}.pkl')
Data.dump(X_test, f'../input/X_test_{fe_name}.pkl')
import numpy as np
import pandas as pd
from ayniy.utils import Data
train = pd.read_csv('../input/train_data.csv')
y_train = train['y']
y_train = np.log1p(y_train)
Data.dump(y_train, '../input/y_train_fe000.pkl')
