def save_feature(df_feat,
prefix,
dir_save,
is_train,
auto_type=True,
list_ignore=[],
is_check=False,
is_viz=True):
DIR_FEATURE = Path('../feature') / dir_save
length = len(df_feat)
if is_check:
for col in df_feat.columns:
if col in list_ignore:
continue
# Nullがあるかどうか
null_len = df_feat[col].dropna().shape[0]
if length - null_len > 0:
print(f"{col} | null shape: {length - null_len}")
# infがあるかどうか
max_val = df_feat[col].max()
min_val = df_feat[col].min()
if max_val == np.inf or min_val == -np.inf:
print(f"{col} | max: {max_val} | min: {min_val}")
print(" * Finish Feature Check.")
sys.exit()
for col in df_feat.columns:
if col in list_ignore:
continue
if auto_type:
feature = df_feat[col].values.astype('float32')
else:
feature = df_feat[col].values
if is_train:
feat_path = DIR_FEATURE / f'{prefix}__{col}_train'
else:
feat_path = DIR_FEATURE / f'{prefix}__{col}_test'
if os.path.exists(str(feat_path) + '.gz'):
continue
else:
if is_viz:
print(f"{feature.shape} | {col}")
utils.to_pkl_gzip(path=str(feat_path), obj=feature)
def ieee_cv(logger, df_train, Y, df_test, COLUMN_GROUP, use_cols, params={}, cols_categorical=[], is_adv=False, is_valid=False):
start_time = "{0:%Y%m%d_%H%M%S}".format(datetime.datetime.now())[:13]
seed = params['seed']
model_type = params['model_type']
n_splits = params['n_splits']
validation = params['fold']
early_stopping_rounds = params['early_stopping_rounds']
del params['seed']
del params['model_type']
del params['n_splits']
del params['fold']
# del params['model_type'], params['n_splits'], params['fold']
if validation=="stratified":
kfold = list(StratifiedKFold(n_splits=n_splits, random_state=seed).split(df_train, Y))
elif validation=='group':
# tmp_kfold = list(GroupKFold(n_splits=n_splits).split(df_train, Y, df_train[COLUMN_GROUP]))
# kfold = [tmp_kfold[3], tmp_kfold[5], tmp_kfold[1], tmp_kfold[4], tmp_kfold[2], tmp_kfold[0]]
kfold = list(GroupKFold(n_splits=n_splits).split(df_train, Y, df_train[COLUMN_GROUP]))
score_list = []
feim_list = []
best_iteration = 0
y_pred = np.zeros(len(df_train))
test_preds = []
if len(df_test):
x_test = df_test
else:
x_test = []
for n_fold, (trn_idx, val_idx) in enumerate(kfold):
# if n_fold!=3:
# continue
x_train = df_train.iloc[trn_idx]
y_train = Y.iloc[trn_idx]
x_valid = df_train.iloc[val_idx]
y_valid = Y.iloc[val_idx]
# if n_fold != 0:
# probing = pd.read_csv('../input/20190929_probing.csv')
# probing = probing[probing['Probing_isFraud']==1]
# probing_ids = probing[COLUMN_ID].values
# y_probing = probing['Probing_isFraud']
# y_probing.name = COLUMN_TARGET
# probing_train = x_test[x_test[COLUMN_ID].isin(probing_ids)]
# print(x_train.shape, y_train.shape)
# x_train = pd.concat([x_train, probing_train], axis=0)
# y_train = pd.concat([y_train, y_probing], axis=0)
# print(x_train.shape, y_train.shape)
x_train = x_train[use_cols]
x_valid = x_valid[use_cols]
val_gr = df_train.iloc[val_idx][COLUMN_GROUP].value_counts()
dtm = val_gr.index.tolist()[0]
print("="*20)
with timer(f" * Fold{n_fold} Validation-{COLUMN_GROUP} {dtm}: {val_gr.values[0]}"):
score, oof_pred, test_pred, feim, best_iter, _ = Classifier(
model_type=model_type,
x_train=x_train,
y_train=y_train,
x_valid=x_valid,
y_valid=y_valid,
x_test=x_test[use_cols],
params=params,
early_stopping_rounds = early_stopping_rounds,
cols_categorical = cols_categorical
)
# if not is_adv:
pb, pv, al = bear_validation(test_pred)
logger.info(f" * Fold{n_fold} {dtm}: {score} | Bear's...PB:{pb} PV:{pv} All:{al}")
print("="*20)
score_list.append(score)
best_iteration += best_iter/n_splits
y_pred[val_idx] = oof_pred
test_preds.append(test_pred)
if len(feim):
feim.rename(columns={'importance': f'imp_fold{n_fold+1}'}, inplace=True)
feim.set_index('feature', inplace=True)
feim_list.append(feim)
cv_score = np.mean(score_list)
cvs = str(cv_score).replace('.', '-')
if len(feim):
df_feim = pd.concat(feim_list, axis=1)
df_feim['imp_avg'] = df_feim.mean(axis=1)
df_feim.sort_values(by='imp_avg', ascending=False, inplace=True)
else:
df_feim = []
## Save
# Each Fold Test Pred
to_pkl_gzip(obj=test_preds, path=f'../output/fold_test_pred/{start_time}_Each_Fold__CV{cvs}__feature{len(use_cols)}')
# Feature Importance
if len(feim):
to_pkl_gzip(obj=df_feim, path=f"../output/feature_importances/{start_time}__CV{cvs}__feature{len(use_cols)}")
#========================================================================
# Adversarial Validationもこちらの関数を使うので、Adversarialの場合はここで終わり
#========================================================================
if is_adv:
pred_result = pd.Series(y_pred, index=df_train[COLUMN_ID].values, name='adv_pred_' + start_time)
return 0, cv_score, df_feim, pred_result, [], []
with timer(" * Make Prediction Result File."):
if is_valid:
pred_result = []
else:
test_pred_avg = np.mean(test_preds, axis=0)
all_pred = np.append(y_pred, test_pred_avg)
all_ids = np.append(df_train[COLUMN_ID].values, df_test[COLUMN_ID].values)
pred_result = pd.DataFrame([all_ids, all_pred], index=[COLUMN_ID, 'pred_' + start_time]).T
pred_result[COLUMN_ID] = pred_result[COLUMN_ID].astype('int')
#========================================================================
# Save
#========================================================================
# Prediction
to_pkl_gzip(obj=pred_result, path=f"../output/pred_result/{start_time}__CV{cvs}__all_preds")
# Submit File
pred_result.columns = [COLUMN_ID, COLUMN_TARGET]
pred_result.iloc[len(df_train):].to_csv(f"../submit/tmp/{start_time}__CV{cvs}__feature{len(use_cols)}.csv", index=False)
return best_iteration, cv_score, df_feim, pred_result, score_list, test_preds
valid_sets=lgb_valid,
early_stopping_rounds=early_stopping_rounds,
num_boost_round=num_boost_round,
verbose_eval=200)
best_iter = estimator.best_iteration
oof_pred = estimator.predict(x_valid)
score = roc_auc_score(y_valid, oof_pred)
cvs = str(score).replace('.', '-')
feim = get_tree_importance(estimator=estimator,
use_cols=x_train.columns)
feim.sort_values(by='importance', ascending=False, inplace=True)
feim['is_valid'] = feim['feature'].map(valid_map)
#========================================================================
# PostProcess
#========================================================================
with timer(" * PostProcess"):
to_pkl_gzip(
obj=feim,
path=
f"../output/selection_feature/{start_time}__CV{cvs}__feature{len(use_cols)}"
)
for path in valid_paths_train:
try:
shutil.move(path, to_dir)
shutil.move(path.replace('train', 'test'), to_dir)
except FileNotFoundError:
print(feature_name)
# # 三行もたないfeatureは各foldをクリアできなかった
# if score < base_fold_score[fold]:
# break
# else:
# cnt +=1
# cv += score/3
df_feim = pd.concat(feim_list, axis=1)
df_feim['imp_avg'] = df_feim.mean(axis=1)
df_feim.sort_values(by='imp_avg', ascending=False, inplace=True)
avg_score = str(np.mean(score_list))[:9].replace('.', '-')
to_pkl_gzip(
obj=df_feim,
path=
f"../output/feature_importances/{start_time}__bear_valid__CV{avg_score}__feature{len(use_cols)}"
)
if cnt == 3:
with open(check_score_path, 'a') as f:
line = f'{feature_name},{cv}\n'
f.write(line)
df_score = pd.read_csv(check_score_path, header=None)
if len(df_score) > 2:
from_dir = 'valid'
to_dir = 'sub_use'
df_score.columns = ['feature', 'score']
df_score.sort_values(by='score', ascending=False, inplace=True)
best_feature = df_score['feature'].values[0]
feim.rename(columns={'importance': f'imp_fold{n_fold+1}'}, inplace=True)
feim.set_index('feature', inplace=True)
feim_list.append(feim)
cv_score = np.mean(score_list)
cvs = str(cv_score).replace('.', '-')
df_feim = pd.concat(feim_list, axis=1)
df_feim['imp_avg'] = df_feim.mean(axis=1)
df_feim.sort_values(by='imp_avg', ascending=False, inplace=True)
## Save
# Feature Importance
to_pkl_gzip(
obj=df_feim,
path=
f"../output/feature_importances/{start_time}__CV{cvs}__{COLUMN_TARGET}__feature{len(use_cols)}"
)
with timer(" * Make Prediction Result File."):
test_pred_avg = np.mean(test_preds, axis=0)
all_pred = np.append(y_pred, test_pred_avg)
all_ids = np.append(tmp_train[COLUMN_ID].values, df_test[COLUMN_ID].values)
pred_result = pd.DataFrame([all_ids, all_pred],
index=[COLUMN_ID, 'pred_' + start_time]).T
pred_result[COLUMN_ID] = pred_result[COLUMN_ID].astype('int')
#========================================================================
# Save
#========================================================================
# Prediction