def lgbm_regressor():
train, test = load_datasets(filename='../input/data_560.csv')
X_train, y_train, X_test, df_coulumns = getDataSet(train, test)
models = ['lr_data_560' + str(i) for i in range(201, 300)]
training_lgbm_regressor(X_train,
y_train,
X_test,
df_coulumns,
ratio=0.8,
model_name=models)
def main():
print('load train test datasets')
train, test = load_datasets(dropDuplicate=False)
batch_size = 1
submit_df = pd.DataFrame({'Id': test['Id']})
submit_pred = np.zeros((test.shape[0], 1))
submit_pred_n = np.zeros((test.shape[0], batch_size))
submit_train = np.zeros((train.shape[0], 1))
submit_train_n = np.zeros((train.shape[0], batch_size))
test_rmses = []
for _, train_all in enumerate(bacth(train, batch_size)):
print('第 %d 批 dataset 开始训练' % _)
y_train_pred, y_test_pred, mean_test_rmse = sub_train(train_all,
test,
hasFilter=False)
submit_pred_n[:, _] = y_test_pred.reshape(-1)
submit_train_n[:, _] = y_train_pred.reshape(-1)
test_rmses.append(mean_test_rmse)
print('第 %d 批 dataset 训练结束' % _)
print('train finished...')
test_rmse = np.mean(test_rmses)
print('mean test rmse: ', test_rmse)
submit_pred[:] = submit_pred_n.mean(1).reshape(-1, 1)
submit_train[:] = submit_train_n.mean(1).reshape(-1, 1)
# test
submit_df['Score_xgb_bagging'] = submit_pred
submission_path_raw = '../models/__models__/{}_{}_{}.csv'.format(
'xgboost_bagging_test', test_rmse,
time.strftime('%m%d%H%M', time.localtime(time.time())))
submission_path_threshold = '../result/{}_{}_{}.csv'.format(
'xgboost_threshold', test_rmse,
time.strftime('%m%d%H%M', time.localtime(time.time())))
submit_df.to_csv(submission_path_raw, index=False)
submit_df = threshold(submit_df, feature='Score_xgb_bagging')
submit_df.to_csv(submission_path_threshold, index=False, header=False)
# train
submit_train_df = train[['Id']]
submit_train_df['Score_xgb_bagging'] = submit_train
submit_train_df.to_csv(submission_path_raw.replace('test', 'train'),
index=False)
print('done.')
def regression():
train, test = load_datasets(fillNan=True)
X_train, y_train, X_test, df_coulumns = getDataSet(train, test)
model_name = ['ridge_1', 'ridge_2', 'lasso_1', 'lasso_2']
clfs = [
Ridge(fit_intercept=True,
alpha=8.858667904100823,
max_iter=500,
normalize=False,
tol=0.01),
Ridge(fit_intercept=True,
alpha=8.858667904100823,
max_iter=500,
normalize=True,
tol=0.01),
Lasso(fit_intercept=True,
alpha=8.858667904100823,
max_iter=500,
normalize=True,
tol=0.01),
Lasso(fit_intercept=True,
alpha=8.858667904100823,
max_iter=500,
normalize=False,
tol=0.01)
]
training_regression(X_train,
y_train,
X_test,
df_coulumns,
clfs,
kBest=True,
k=476,
ratio=1,
model_name=model_name)
model_name = ['ridge_3', 'ridge_4', 'lasso_3', 'lasso_4']
training_regression(X_train,
y_train,
X_test,
df_coulumns,
clfs,
minMaxScaler=MinMaxScaler((-1, 1)),
kBest=True,
k=476,
ratio=1,
model_name=model_name)
def main():
print('load train test datasets')
usebatch = False
# configuration display
print('batch is used: {}'.format("Yes" if usebatch else "No"))
with open('../input/feature.info', 'a+') as outf:
if usebatch:
for train, test in batch_with_gain_importance(featurefile='../models/info/gain_importance_data.csv', filename='../input/data.csv'):
_train(train, test, outf)
else:
train, test = load_datasets(filename='../input/data_560.csv')
_train(train, test, outf)
def main():
# 加载 数据集
train, text = load_datasets()
# 加载 feature importance
features = most_importance('split_importance_02181349.csv')
best_feature = []
maxRmse = 0
drop_feature = []
drop_len = 10
drop_freq = defaultdict(int)
all_feature = features[::]
while len(features) != 0:
choose_feature = features.pop(0)
candicate_feature = best_feature.copy()
candicate_feature.append(choose_feature)
train_feature = train[candicate_feature]
train_label = train['Score']
train_feature, valid_feature, train_label, valid_label = train_test_split(
train_feature, train_label, test_size=0.3, random_state=0)
train_feature = np.array(train_feature.values)
valid_feature = np.array(valid_feature.values)
train_label = np.array(train_label).reshape(-1)
valid_label = np.array(valid_label).reshape(-1)
lgb_train = lgbm.Dataset(train_feature, label=train_label)
lgb_eval = lgbm.Dataset(valid_feature,
label=valid_label,
reference=lgb_train)
lgbm_params = {
'boosting_type': 'gbdt',
'objective': 'regression',
'metric': 'rmse',
'min_child_weight': 20,
'num_leaves': 2**5,
'lambda_l2': 2,
'subsample': 0.5,
'colsample_bytree': 1,
'learning_rate': 0.1,
'seed': 2017,
'verbose': 100,
'silent': True,
}
model = lgbm.train(lgbm_params,
lgb_train,
num_boost_round=5000,
valid_sets=lgb_eval,
early_stopping_rounds=100)
valid_pred = model.predict(valid_feature,
num_iteration=model.best_iteration)
valid_rmse = rmse(valid_label, valid_pred)
valid_rmse = 1 / (1 + valid_rmse)
if valid_rmse >= maxRmse - 0.000245:
best_feature.append(choose_feature)
maxRmse = max(maxRmse, valid_rmse)
with open('../models/info/sub_feature_02182137.csv', 'a+') as out:
out.write(
str(valid_rmse) + ',' + choose_feature + ',' +
str(len(drop_feature)) + '\n')
else:
drop_freq[choose_feature] += 1
drop_feature.append(choose_feature)
if len(drop_feature) == drop_len:
for val in drop_feature[::-1]:
if drop_freq[val] > 1: continue
features.insert(0, val)
drop_feature = []
drop_len += drop_len
with open('../models/info/sub_feature_02182137.log', 'a+') as out:
out.write('当前 drop 队列长度 {}：'.format(len(drop_feature)) + '\n')
out.write(' || '.join(drop_feature) + '\n')
for drop in drop_feature:
out.write('{} {}'.format(drop_freq[drop], drop) + '\n')
def main():
print('load train test datasets')
train, test = load_datasets()
submit_df = pd.DataFrame({'userid': test['Id']})
X_train = train.drop(['Id', 'Score'], axis=1)
X_test = test.drop(['Id'], axis=1)
y_train = train['Score'] - 1
df_columns = X_train.columns
xgb_params = {
'eta': 0.01,
'min_child_weight': 20,
'colsample_bytree': 0.5,
'max_depth': 10,
'subsample': 0.5,
'lambda': 2.0,
'scale_pos_weight': 1,
'eval_metric': 'mlogloss',
'objective': 'multi:softmax',
'silent': 1,
'booster': 'gbtree',
'num_class': 5
}
dtrain_all = xgb.DMatrix(X_train.values,
y_train.values,
feature_names=df_columns)
dtest = xgb.DMatrix(X_test.values, feature_names=df_columns)
# 5 折交叉验证
nfold = 5
cv_result = xgb.cv(dict(xgb_params),
dtrain_all,
nfold=nfold,
stratified=True,
num_boost_round=10000,
early_stopping_rounds=100,
verbose_eval=100,
show_stdv=False)
best_num_boost_rounds = len(cv_result)
mean_train_mlogloss = cv_result.loc[best_num_boost_rounds -
11:best_num_boost_rounds - 1,
'train-mlogloss-mean'].mean()
mean_test_mlogloss = cv_result.loc[best_num_boost_rounds -
11:best_num_boost_rounds - 1,
'test-mlogloss-mean'].mean()
print('best_num_boost_rounds = {}'.format(best_num_boost_rounds))
# num_boost_round = int(best_num_boost_rounds * 1.1)
# print('num_boost_round = ', num_boost_round)
print('mean_rmse_auc = {:.7f} , mean_rmse_auc = {:.7f}\n'.format(
mean_train_mlogloss, mean_test_mlogloss))
print('---> training on total dataset')
model = xgb.train(dict(xgb_params),
dtrain_all,
num_boost_round=best_num_boost_rounds)
print('---> predict test')
y_pred = model.predict(dtest, ntree_limit=model.best_ntree_limit)
submit_df['Score'] = y_pred
submit_df['Score'] = submit_df['Score'] + 1
print(y_pred)
submission_path = '../result/{}_{}.csv'.format('xgb', mean_test_mlogloss)
submit_df.to_csv(submission_path, index=False, header=False)
print('done.')