def gen_sub_by_para(drop_useless_pkg, drop_long):
args = locals()
lda_feature = get_lda_from_usage()
feature = extend_feature(span_no=24,
input=lda_feature,
drop_useless_pkg=False,
drop_long=False)
feature = extend_device_brand(feature)
feature_label = attach_device_train_label(feature)
train = feature_label[feature_label['sex'].notnull()]
test = feature_label[feature_label['sex'].isnull()]
X = train.drop(['sex', 'age', 'sex_age', 'device'], axis=1)
Y = train['age']
Y_CAT = pd.Categorical(Y)
X_train, X_test, y_train, y_test = train_test_split(X, Y_CAT.labels)
lgb_train = lgb.Dataset(X_train, label=y_train)
lgb_eval = lgb.Dataset(X_test, y_test, reference=lgb_train)
params = {
'boosting_type': 'gbdt',
'max_depth': 3,
'random_state': 47,
# "min_data_in_leaf":1000,
'verbose': -1,
'colsample_bytree': 0.58,
# 'min_child_samples': 289,
# 'min_child_weight': 0.1,
'min_data_in_leaf': 1472,
# 'num_leaves': 300,
'reg_alpha': 3,
'reg_lambda': 4,
'subsample': 0.8
}
params_age = {
'metric': {'multi_logloss'},
'num_class': 11,
'objective': 'multiclass',
}
params_sex = {
'metric': ['auc', 'binary_logloss'],
'objective': 'binary',
}
try:
gbm = lgb.train(dict(params, **params_age),
lgb_train,
num_boost_round=1000,
valid_sets=lgb_eval,
early_stopping_rounds=50)
except Exception as error:
print(
f'Model input columns:{list(X.columns)}\n dict({X.dtypes.sort_values()})'
)
raise error
best = round(gbm.best_score.get('valid_0').get('multi_logloss'), 5)
best = "{:.5f}".format(best)
pre_x = test.drop(['sex', 'age', 'sex_age', 'device'], axis=1)
sub = pd.DataFrame(
gbm.predict(pre_x.values, num_iteration=gbm.best_iteration))
# sub[1] = 1 - sub[0]
# sub.head(3)
sub.columns = Y_CAT.categories
print(sub.columns)
sub['DeviceID'] = test['device'].values
sub = sub[[
'DeviceID', '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '10'
]]
sub.head(3)
#lgb.plot_importance(gbm, max_num_features=20)
print(
f'=============Final train feature({len(feature_label.columns)}):\n{list(feature_label.columns)} \n {len(feature_label.columns)}'
)
file = f'./sub/baseline_age_{best}.csv'
#print(f'sub file save to {file}')
sub.to_csv(file, index=False)
def gen_sub_by_para():
args = locals()
logger.debug(f'Run train dnn:{args}')
# feature_label = get_dynamic_feature(svd_cmp)
feature_label = get_stable_feature('1006')
train=feature_label[feature_label['sex'].notnull()]
test =feature_label[feature_label['sex'].isnull()]
X = train.drop(['sex', 'age', 'sex_age', 'device'], axis=1)
Y = train['sex_age']
Y_CAT = pd.Categorical(Y)
X_train, X_test, y_train, y_test = train_test_split(X, Y_CAT.labels)
lgb_train = lgb.Dataset(X_train, label=y_train)
lgb_eval = lgb.Dataset(X_test, y_test, reference=lgb_train)
params = {
'boosting_type': 'gbdt',
'metric': {'multi_logloss'},
'num_class': 22,
'objective': 'multiclass',
'random_state': 47,
'verbose': -1,
#'num_leaves':num_leaves,
'max_depth': 3,
#"min_data_in_leaf":1000,
'feature_fraction': 0.2,
'subsample': 0.4,
'reg_alpha': 6,
'reg_lambda': 4,
# 'min_child_samples': 289,
#'min_child_weight': 0.1,
'min_data_in_leaf': 1472,
#'num_leaves': 300,
'learning_rate' : 0.01#,
}
try:
gbm = lgb.train(params,
lgb_train,
num_boost_round=20000,
valid_sets=[lgb_train, lgb_eval,],
early_stopping_rounds=50,)
print(f"Light GBM:{str(gbm)}")
except Exception as error:
print(f'Model input columns:{list(X.columns)}\n dict({X.dtypes.sort_values()})')
raise error
best_score = round(gbm.best_score.get('valid_1').get('multi_logloss'), 5)
best_epoch = gbm.best_iteration
pre_x=test.drop(['sex','age','sex_age','device'],axis=1)
# sub=pd.DataFrame(gbm.predict(pre_x.values,num_iteration=gbm.best_iteration))
#
# sub.columns=Y_CAT.categories
# sub['DeviceID']=test['device'].values
# sub=sub[['DeviceID', '1-0', '1-1', '1-2', '1-3', '1-4', '1-5', '1-6', '1-7','1-8', '1-9', '1-10', '2-0', '2-1', '2-2', '2-3', '2-4', '2-5', '2-6', '2-7', '2-8', '2-9', '2-10']]
# from sklearn.metrics import log_loss
# loss_best = log_loss(y_test, gbm.predict(X_test,num_iteration=gbm.best_iteration))
#
# loss_default = log_loss(y_test, gbm.predict(X_test))
#
# print(f'Loss_best={loss_best}, Loss_default={loss_default}')
# #lgb.plot_importance(gbm, max_num_features=20)
#
# # print(f'=============Final train feature({len(feature_label.columns)}):\n{list(feature_label.columns)} \n {len(feature_label.columns)}')
#
# file = f'./sub/baseline_lg_{best_score}_{best_epoch}_{args}.csv'
# file = replace_invalid_filename_char(file)
# print(f'sub file save to {file}')
# sub.to_csv(file,index=False)
train_bk = pd.DataFrame(gbm.predict(train.drop(['sex', 'age', 'sex_age', 'device'], axis=1)),
index=train.device,
columns=Y_CAT.categories
)
test_bk = pd.DataFrame(gbm.predict(pre_x),
index=test.device,
columns=Y_CAT.categories
)
from code_felix.tiny.util import save_result_for_ensemble
save_result_for_ensemble(f'{best_score}_{best_epoch}_lgb_{args}',
train=train_bk,
test=test_bk,
label=None,
)
def gen_sub_by_para(drop_feature):
args = locals()
logger.debug(f'Run train dnn:{args}')
#feature_label = get_dynamic_feature(None)
feature_label = get_stable_feature('1011')
#feature_label = random_feature(feature_label, 1/2)
feature_label = get_cut_feature(feature_label, drop_feature)
#feature_label = get_best_feautre(feature_label)
# daily_info = summary_daily_usage()
# feature_label = feature_label.merge(daily_info, on='device', how='left')
train = feature_label[feature_label['sex'].notnull()]
test = feature_label[feature_label['sex'].isnull()]
X = train.drop(['sex', 'age', 'sex_age', 'device'], axis=1)
Y = train['sex_age']
Y_CAT = pd.Categorical(Y)
X_train, X_test, y_train, y_test = train_test_split(X, Y_CAT.codes)
gbm = get_model()
logger.debug(f"Run the xgb with:{gpu_params}")
# print(random_search.grid_scores_)
gbm.fit(X, Y_CAT.codes, verbose=True)
#results = gbm.evals_result()
#print(results)
#
# best_epoch = np.array(results['validation_0']['mlogloss']).argmin() + 1
# best_score = np.array(results['validation_1']['mlogloss']).min()
#
pre_x = test.drop(['sex', 'age', 'sex_age', 'device'], axis=1)
print_imp_list(X_train, gbm)
###Save result for ensemble
train_bk = pd.DataFrame(gbm.predict_proba(
train.drop(['sex', 'age', 'sex_age', 'device'], axis=1)),
index=train.device,
columns=Y_CAT.categories)
test_bk = pd.DataFrame(gbm.predict_proba(pre_x),
index=test.device,
columns=Y_CAT.categories)
label_bk = pd.DataFrame(
{'label': Y_CAT.codes},
index=train.device,
)
save_result_for_ensemble(
f'all_xgb_col_{len(feature_label.columns)}_{args}',
train=train_bk,
test=test_bk,
label=label_bk,
)
def gen_sub_by_para():
args = locals()
logger.debug(f'Run train dnn:{args}')
#feature_label = get_dynamic_feature(None)
feature_label = get_stable_feature('1011')
#feature_label = get_cut_feature(feature_label, drop_feature)
train = feature_label[feature_label['sex'].notnull()]
test = feature_label[feature_label['sex'].isnull()]
X = train.drop(['sex', 'age', 'sex_age', 'device'], axis=1)
Y = train['sex_age']
Y_CAT = pd.Categorical(Y)
X_train, X_test, y_train, y_test = train_test_split(X, Y_CAT.codes)
gbm = get_model()
logger.debug(f"Run the xgb with:{gpu_params}")
# print(random_search.grid_scores_)
gbm.fit(X_train,
y_train,
eval_set=[(X_train, y_train), (X_test, y_test)],
early_stopping_rounds=50,
verbose=True)
results = gbm.evals_result()
logger.debug(results)
best_epoch = np.array(results['validation_1']['mlogloss']).argmin() + 1
best_score = np.array(results['validation_1']['mlogloss']).min()
logger.debug(f"Xgb arrive {best_score} at {best_epoch}")
pre_x = test.drop(['sex', 'age', 'sex_age', 'device'], axis=1)
print_imp_list(X_train, gbm)
###Save result for ensemble
train_bk = pd.DataFrame(gbm.predict_proba(
train.drop(['sex', 'age', 'sex_age', 'device'], axis=1)),
index=train.device,
columns=get_category().categories)
test_bk = pd.DataFrame(gbm.predict_proba(pre_x),
index=test.device,
columns=get_category().categories)
label_bk = pd.DataFrame(
{'label': Y_CAT.codes},
index=train.device,
)
save_result_for_ensemble(
f'all_xgb_{args}',
train=train_bk,
test=test_bk,
label=label_bk,
)
for file in file_list:
train, label, test = read_result_for_ensemble(file)
train_list.append(train)
if label is not None: label_list.append(label)
test_list.append(test)
train = pd.concat(train_list, axis=1)
test = pd.concat(test_list, axis=1)
label = label_list[0]
train = train.sort_index()
label = label.sort_index()
X_train, X_test, y_train, y_test = train_test_split(train, label.iloc[:,0])
#drop_list = list(np.arange(0.65, 0.7, 0.03))
# drop_list.reverse()
for dense in [128]:
for drop_out in [0.63]:
drop_out = round(drop_out, 2)
patience=50
lr = 0.0005
#搭建融合后的模型
inputs = Input((X_train.shape[1:]))
x = Dropout(drop_out)(inputs)
x = Dense(dense, activation='relu')(x)