def load_data_from_pickle(self):
print('Loading Preprocessed Data...')
x_train = utils.load_pkl_to_data(self.preprocessed_path + 'x_train_gan.p')
x_test = utils.load_pkl_to_data(self.preprocessed_path + 'x_test_gan.p')
return x_train, x_test
def train(self, load_pickle=False, load_pickle_path=None):
"""
Training the model
"""
start_time = time.time()
path_list = [
self.pred_path + 'multiclass/', self.pred_path + 'pred_era/',
self.pred_path + 'pred_era/final_results/',
self.pred_path + 'final_results/'
]
utils.check_dir(path_list)
print('======================================================')
print('Start Training PrejudgeMultiClass...')
if load_pickle:
# Load era_sign_test
if load_pickle_path is None:
era_sign_test = utils.load_pkl_to_data(
self.prejudged_data_path + 'multiclass_era_sign_test.p')
else:
era_sign_test = utils.load_pkl_to_data(load_pickle_path)
else:
# Training Era Sign
era_sign_test = self.predict_era_sign()
# era_sign_test = self.load_era_sign_csv(self.pred_path + 'pred_era/final_results/lgb_result.csv')
# Save era_sign_test to Pickle File
utils.save_data_to_pkl(
era_sign_test,
self.prejudged_data_path + 'multiclass_era_sign_test.p')
# Get Split Data
x_test, x_g_test, id_test, x_test_idx = self.split_test_set_by_era_sign(
era_sign_test)
x_train, x_g_train, y_train, w_train, e_train = self.split_train_set_by_era(
)
# Training Models by Era Sign
prob_test = \
self.train_models_by_era_sign(x_train, x_g_train, y_train, w_train, e_train,
x_test, x_g_test, id_test, x_test_idx)
# Save Predictions
utils.save_pred_to_csv(self.pred_path + 'final_results/prejudge_',
self.id_test, prob_test)
total_time = time.time() - start_time
print('======================================================')
print('Training Done!')
print('Total Time: {}s'.format(total_time))
print('======================================================')
def random_split(x, y, n_splits=None, n_cv=None, cv_seed=None):
train_data = utils.load_pkl_to_data(
'./data/preprocessed_data/x_g_train.p')
data_mt = np.array(train_data)
index = data_mt[:, 2]
# station_list = index.tolist()
# min_number = 10000
# for i in np.unique(index):
# if min_number > station_list.count(i):
# min_number = station_list.count(i)
# if n_splits > min_number:
# raise ValueError(
# '--The least populated station has only %d members,please input new cv_number--' % min_number)
cv_count = 0
skf = StratifiedKFold(n_splits=n_cv,
shuffle=True,
random_state=cv_seed)
for train_index, valid_index in skf.split(index, index):
# Training data
x_train = x[train_index]
y_train = y[train_index]
# Validation data
x_valid = x[valid_index]
y_valid = y[valid_index]
cv_count += 1
utils.print_cv_info(cv_count, n_cv)
yield x_train, y_train, x_valid, y_valid
def __init__(self, seed):
utils.check_dir([tampered_pred_path])
np.random.seed(seed)
base_result = pd.read_csv(base_fake_result_path + '0.5_fake_result.csv', header=0, dtype=np.float64)
self.prob = np.array(base_result['proba'], dtype=np.float64)
self.id_test = utils.load_pkl_to_data(preprocessed_data_path + 'id_test.p')
self.same_idx_list = utils.load_pkl_to_data(preprocessed_data_path + 'same_test_idx_pairs.p')
self.code_id_train, self.code_id_test = utils.load_preprocessed_code_id(preprocessed_data_path)
same_test_df = pd.read_csv(preprocessed_data_path + 'same_test_pairs.csv', header=0, dtype=np.float64)
self.same_test_code_id = same_test_df['code_id']
self.same_test_id = same_test_df['id']
self.test_id_to_idx_dict = {}
for idx, id_ in enumerate(self.id_test):
self.test_id_to_idx_dict[id_] = idx
def generate_tampered_results_by_weight(self, n_pairs):
w_train = utils.load_pkl_to_data(preprocessed_data_path + 'w_train.p')
print('------------------------------------------------------')
print('Calculating Big Weight Same Pairs...')
print('------------------------------------------------------')
print('Sorting...')
sorted_by_weight_idx = np.argsort(w_train)[:-n_pairs*3:-1]
sorted_w_train = w_train[sorted_by_weight_idx]
sorted_code_id_train = self.code_id_train[sorted_by_weight_idx]
print('Deduplicating...')
big_weight_code_id = []
big_weight_w_train = []
for idx, code_id in enumerate(sorted_code_id_train):
if code_id not in big_weight_code_id:
if code_id in set(self.same_test_code_id):
big_weight_code_id.append(code_id)
big_weight_w_train.append(sorted_w_train[idx])
print('Generating Pairs...')
idx_pair_list, w_train_col, code_id_col, id_col = \
self.get_pair_list(n_pairs, big_weight_code_id, use_weight=True, w_train_list=big_weight_w_train)
print('------------------------------------------------------')
print('Number of Big Weight Same Pairs: {}'.format(len(idx_pair_list)))
utils.save_data_to_pkl(idx_pair_list, preprocessed_data_path + 'big_weight_idx_pairs.p')
index = []
for i in range(1, len(idx_pair_list)+1):
index.extend([i, i])
df_log = pd.DataFrame({'index': np.array(index, dtype=int),
'weight': np.array(w_train_col),
'code_id': np.array(code_id_col, dtype=int),
'id': np.array(id_col, dtype=int)})
cols = ['index', 'weight', 'code_id', 'id']
df_log = df_log.loc[:, cols]
tampered_pred_path_ = tampered_pred_path + 'big_weight_tampered_log.csv'
print('------------------------------------------------------')
print('Saving {} ...'.format(tampered_pred_path_))
df_log.to_csv(tampered_pred_path_, sep=',', index=False)
# Save Same Pairs csv file
self.save_same_pairs_test_csv(preprocessed_data_path + 'big_weight_same_pairs.csv', idx_pair_list)
# Generate Tampered Results
self.tamper_result(idx_pair_list, 'big_weight')
def generate_tampered_results_by_absence(self, n_pairs):
diff_code_id_test = utils.load_pkl_to_data(preprocessed_data_path + 'diff_code_id_test.p')
print('------------------------------------------------------')
print('Calculating Absent Same Pairs...')
print('------------------------------------------------------')
print('Sorting...')
diff_code_id_test = np.sort(diff_code_id_test)
absent_code_id = diff_code_id_test[:-n_pairs*3:-1]
print('Generating Pairs...')
idx_pair_list, code_id_col, id_col = self.get_pair_list(n_pairs, absent_code_id)
print('------------------------------------------------------')
print('Number of Absent Same Pairs: {}'.format(len(idx_pair_list)))
utils.save_data_to_pkl(idx_pair_list, preprocessed_data_path + 'absent_idx_pairs.p')
index = []
for i in range(1, len(idx_pair_list)+1):
index.extend([i, i])
df_log = pd.DataFrame({'index': np.array(index, dtype=int),
'code_id': np.array(code_id_col, dtype=int),
'id': np.array(id_col, dtype=int)})
cols = ['index', 'code_id', 'id']
df_log = df_log.loc[:, cols]
tampered_pred_path_ = tampered_pred_path + 'absent_tampered_log.csv'
print('------------------------------------------------------')
print('Saving {} ...'.format(tampered_pred_path_))
df_log.to_csv(tampered_pred_path_, sep=',', index=False)
# Save Same Pairs csv file
self.save_same_pairs_test_csv(preprocessed_data_path + 'absent_same_pairs.csv', idx_pair_list)
# Generate Tampered Results
self.tamper_result(idx_pair_list, 'absent')
def train(self, load_pickle=False, load_pickle_path=None):
"""
Training the model
"""
start_time = time.time()
path_list = [
self.pred_path + 'positive/', self.pred_path + 'negative/',
self.pred_path + 'pred_era/',
self.pred_path + 'pred_era/final_results/',
self.pred_path + 'final_results/',
self.loss_log_path + 'positive/', self.loss_log_path + 'negative/'
]
utils.check_dir(path_list)
print('======================================================')
print('Start Training PrejudgeBinary...')
if load_pickle:
# Load era_sign_test
if load_pickle_path is None:
era_sign_test = utils.load_pkl_to_data(
self.prejudged_data_path + 'binary_era_sign_test.p')
else:
era_sign_test = utils.load_pkl_to_data(load_pickle_path)
else:
# Training Era Sign
era_sign_test = self.predict_era_sign()
# era_sign_test = self.load_era_sign_csv(self.pred_path + 'pred_era/final_results/lgb_result.csv')
# Save era_sign_test to Pickle File
utils.save_data_to_pkl(
era_sign_test,
self.prejudged_data_path + 'binary_era_sign_test.p')
# Print Prediction of Positive Era Rate
utils.print_positive_rate_test(era_sign_test)
# Get Split Data
x_test_p, x_g_test_p, id_test_p, era_idx_test_p, x_test_n, \
x_g_test_n, id_test_n, era_idx_test_n = self.split_test_set_by_era_sign(era_sign_test)
# Training Models by Era Sign
prob_test = \
self.train_models_by_era_sign(x_test_p, x_g_test_p, id_test_p, era_idx_test_p,
x_test_n, x_g_test_n, id_test_n, era_idx_test_n)
# Save Predictions
utils.save_pred_to_csv(self.pred_path + 'final_results/prejudge_',
self.id_test, prob_test)
# Print Prediction of Positive Era Rate
utils.print_positive_rate_test(era_sign_test)
total_time = time.time() - start_time
print('======================================================')
print('Training Done!')
print('Total Time: {}s'.format(total_time))
print('======================================================')
def split_data_by_gan(self,
load_pickle=True,
sample_ratio=None,
sample_by_era=True,
generate_mode='valid'):
print('======================================================')
print('Splitting Adversarial Validation Set by GAN...')
if load_pickle:
similarity_prob = utils.load_pkl_to_data(cfg.gan_prob_path +
'similarity_prob.p')
else:
similarity_prob = \
GenerateValidation.train(train_path=cfg.train_csv_path, test_path=cfg.test_csv_path, global_epochs=1,
similarity_prob_path=cfg.gan_prob_path, return_similarity_prob=True,
load_preprocessed_data=True)
valid_idx = []
train_idx = []
if sample_by_era:
similarity_prob_e = []
index_e = []
similarity_prob_all = []
index_all = []
era_tag = 0
era_all = [era_tag]
for idx, era in enumerate(self.e_train):
if idx == len(self.e_train) - 1:
similarity_prob_e.append(similarity_prob[idx])
index_e.append(idx)
similarity_prob_all.append(similarity_prob_e)
index_all.append(index_e)
elif era_tag == era:
similarity_prob_e.append(similarity_prob[idx])
index_e.append(idx)
else:
era_tag = era
era_all.append(era)
similarity_prob_all.append(similarity_prob_e)
index_all.append(index_e)
similarity_prob_e = [similarity_prob[idx]]
index_e = [idx]
for e, similarity_prob_e in enumerate(similarity_prob_all):
n_sample_e = int(len(similarity_prob_e) * sample_ratio)
most_similar_idx_e = np.argsort(
similarity_prob_e)[:, :-(n_sample_e + 1):-1]
least_similar_idx_e = np.argsort(
similarity_prob_e)[:, :len(similarity_prob_e) - n_sample_e]
if generate_mode == 'valid':
valid_idx += list(index_all[e][most_similar_idx_e])
train_idx += list(index_all[e][least_similar_idx_e])
elif generate_mode == 'train':
train_idx += list(index_all[e][most_similar_idx_e])
valid_idx += list(index_all[e][least_similar_idx_e])
else:
raise ValueError("Wrong 'generate_mode'!")
else:
n_sample = int(len(similarity_prob) * sample_ratio)
most_similar_idx = np.argsort(similarity_prob)[:, :-(n_sample +
1):-1]
least_similar_idx = np.argsort(
similarity_prob)[:, :len(similarity_prob) - n_sample]
if generate_mode == 'valid':
valid_idx = most_similar_idx
train_idx = least_similar_idx
elif generate_mode == 'train':
train_idx = least_similar_idx
valid_idx = most_similar_idx
else:
raise ValueError("Wrong 'generate_mode'!")
# Generate Validation Set
self.x_valid = self.x_train[valid_idx]
self.y_valid = self.x_train[valid_idx]
# Generate Training Set
self.x_train = self.x_train[train_idx]
self.y_train = self.y_train[train_idx]
self.w_train = self.w_train[train_idx]
self.e_train = self.e_train[train_idx]
if group_list is not None:
self.x_g_valid = self.x_g_train[valid_idx]
self.x_g_train = self.x_g_train[train_idx]
# Save Adversarial Validation Set
print('Saving Adversarial Validation Set...')
utils.save_data_to_pkl(self.x_valid,
self.preprocess_path + 'x_valid.p')
utils.save_data_to_pkl(self.x_g_valid,
self.preprocess_path + 'x_g_valid.p')
utils.save_data_to_pkl(self.y_valid,
self.preprocess_path + 'y_valid.p')
