def model_train_done_rnn(model_id, lag_dates, pcas, file_path_list,
sub_model_names, outputs_list, events_set, model_dir):
"""
更新t_event_model_file、t_event_model_detail、t_event_model_tran表的数据。
:param model_id: string. 模型编号
:param file_path_list: array. 模型地址
"""
db = DatabaseWrapper(dbname=event_dbname)
try:
db.begin_transaction()
sql = "INSERT INTO t_event_model_file(file_id, file_url, model_id) VALUES (%s, %s, %s) "
params = []
for model_fp in file_path_list:
param = (UuidHelper.guid(), model_fp, model_id)
params.append(param)
db.executemany(sql, params)
# 子模型信息入库
detail_ids = []
sql = "INSERT INTO t_event_model_detail(detail_id, model_name, status, model_id, lag_date, pca, create_date, " \
"create_time) values (%s, %s, %s, %s, %s, %s, %s, %s)"
params = []
for sub_model_name, lag_date, pca in zip(sub_model_names, lag_dates,
pcas):
detail_id = UuidHelper.guid()
detail_ids.append(detail_id)
params.append(
(detail_id, sub_model_name, DataStatus.SUCCESS.value, model_id,
int(lag_date), int(pca), sys_date(sys_date_formatter),
sys_time(sys_time_formatter)))
db.executemany(sql, params)
# 分事件模型信息入库
sql = "INSERT INTO t_event_model_tran(tran_id, event_name, num, detail_id, status, create_date, create_time) " \
"values (%s, %s, %s, %s, %s, %s, %s)"
params = []
for detail_id, outputs in zip(detail_ids, outputs_list):
events_num = pp.get_event_num(outputs, events_set)
for e in events_set:
tran_id = UuidHelper.guid()
event_num = events_num[e]
param = (tran_id, e, event_num,
detail_id, DataStatus.SUCCESS.value,
sys_date(sys_date_formatter),
sys_time(sys_time_formatter))
params.append(param)
db.executemany(sql, params)
sql = "UPDATE t_event_model SET model_dir = %s WHERE model_id = %s"
db.execute(sql, (model_dir, model_id))
db.commit()
return detail_ids
except Exception as e:
db.rollback()
raise RuntimeError(e)
finally:
db.close()
def evaluate(self):
inputs_test, outputs_test = \
preprocess.gen_samples_by_pred_date(self.data,
self.events_p_oh,
self.input_len,
self.output_len,
self.dates,
self.eval_start_date,
self.eval_end_date)
event_col = self.events_set.index(self.eval_event)
outputs_test = outputs_test[:, :, event_col]
outputs_test = np.reshape(outputs_test, [*outputs_test.shape, 1])
events_num = preprocess.get_event_num(outputs_test, [self.eval_event])
preds = predict_cnn.predict_samples(self.model, inputs_test)
evals_summary, evals_separate, eval_events = \
model_evalution.evaluate_sub_model(
preds,
outputs_test,
[self.eval_event],
events_num)
return evals_summary[0]
def evaluate(self):
inputs_test, outputs_test = \
preprocess.gen_samples_by_pred_date(self.data,
self.events_p_oh,
self.input_len,
self.output_len,
self.dates,
self.eval_start_date,
self.eval_end_date)
events_num = preprocess.get_event_num(outputs_test, self.events_set)
preds = predict_rnn.predict_samples(self.encoder, self.decoder,
inputs_test, self.output_len,
len(self.events_set))
evals_summary, evals_separate, eval_events = \
model_evalution.evaluate_sub_model(
preds,
outputs_test,
self.events_set,
events_num)
return evals_summary[0]
def model_train_done_cnn(model_id, kernel_size_array, pool_size_array,
lag_date_array, file_path_array, sub_model_name_array,
output_array, event_set_array, model_dir):
"""
更新t_event_model_file、t_event_model_detail、t_event_model_tran表的数据。
:param model_id: string. 模型编号
:param kernel_size_array: array. 卷积核列表
:param pool_size_array: array. 过滤器列表
:param lag_date_array: array. 滞后期列表
:param file_path_array: array. 模型文件地址列表
:param sub_model_name_array: array. 模型文件名列表
:param output_array: array.事件样本列表（事件表数据）
:param event_set_array: array.事件类别列表
:param model_dir: str.模型存放地址
:return array, 子模型编号列表
"""
db = DatabaseWrapper(dbname=event_dbname)
try:
db.begin_transaction()
sql = "INSERT INTO t_event_model_file(file_id, file_url, model_id) VALUES (%s, %s, %s) "
params = []
for model_fp in file_path_array:
param = (UuidHelper.guid(), model_fp, model_id)
params.append(param)
db.executemany(sql, params)
# 子模型信息入库
detail_ids = []
sql = "INSERT INTO t_event_model_detail(detail_id, model_name, status, model_id, lag_date, kernel_size, " \
"pool_size, create_date, create_time) values (%s, %s, %s, %s, %s, %s, %s, %s, %s)"
params = []
for sub_model_name, lag_date, kernel_size, pool_size in zip(
sub_model_name_array, lag_date_array, kernel_size_array,
pool_size_array):
detail_id = UuidHelper.guid()
detail_ids.append(detail_id)
params.append(
(detail_id, sub_model_name, DataStatus.SUCCESS.value, model_id,
int(lag_date), int(kernel_size), int(pool_size),
sys_date(sys_date_formatter), sys_time(sys_time_formatter)))
db.executemany(sql, params)
# 分事件模型信息入库
sql = "INSERT INTO t_event_model_tran(tran_id, event_name, num, detail_id, status, create_date, create_time) " \
"values (%s, %s, %s, %s, %s, %s, %s)"
params = []
for detail_id, outputs in zip(detail_ids, output_array):
events_num = pp.get_event_num(outputs, event_set_array)
for event in event_set_array:
tran_id = UuidHelper.guid()
event_num = events_num[event]
param = (tran_id, event, event_num,
detail_id, DataStatus.SUCCESS.value,
sys_date(sys_date_formatter),
sys_time(sys_time_formatter))
params.append(param)
db.executemany(sql, params)
sql = "UPDATE t_event_model SET model_dir = %s WHERE model_id = %s"
db.execute(sql, (model_dir, model_id))
db.commit()
return detail_ids
except Exception as e:
db.rollback()
raise RuntimeError(e)
finally:
db.close()
def evaluate_models(data, dates, detail_ids, sub_model_dirs, params_list,
events_p_oh, events_set, n_classes, start_date, end_date,
model_type, eval_event, model_dir):
"""
评估子模型并将评估结果存入数据库
Args:
data: 评估范围内的数据表数据, 用于生成预测值
dates: 数据表的日期列表, 全部日期, 不仅为评估范围内
detail_ids: 子模型对应的 detail_id 列表
sub_model_dirs: 子模型对应的模型文件存放路径列表
params_list: 子模型对应的超参数列表
events_p_oh: 补 0 one-hot 形式的事件列表, 全部日期范围
events_set: 去重后有序事件集合
n_classes: 事件类别个数, 包含 0 事件, 为预测值最后一维长度
start_date: 开始评估日期, 为预测开始日期
end_date: 结束评估日期, 为预测结束日期
model_type: string.模型类型，对应代码项为ModelType
eval_event: cnn模型专用参数，该参数不为none时表示该模型是cnn
model_dir: str，模型存放地址
Returns:
scores: 子模型综合评分列表
events_num: dict, 测试集事件出现次数: {event: event_num}
"""
scores = []
events_num = {}
for detail_id, sub_model_dir, params in zip(detail_ids, sub_model_dirs,
params_list):
logger.info(f'评估模型: {sub_model_dir}, detail_id: {detail_id}')
if ModelType.CNN.value == model_type:
input_len, output_len = params
inputs_test, outputs_test = pp.gen_samples_by_pred_date(
data, events_p_oh, input_len, output_len, dates, start_date,
end_date)
event_col = events_set.index(eval_event)
outputs_test = outputs_test[:, :, event_col]
outputs_test = np.reshape(outputs_test, [*outputs_test.shape, 1])
events_set_ = [eval_event]
model = load_cnn_model(sub_model_dir)
preds = predict_cnn.predict_samples(model, inputs_test)
elif ModelType.RNN.value == model_type:
input_len, output_len, n_pca = params
values_pca = pp.apply_pca(n_pca, model_dir, data, True)
events_set_ = events_set
inputs_test, outputs_test = \
pp.gen_samples_by_pred_date(values_pca, events_p_oh, input_len, output_len, dates, start_date, end_date)
encoder, decoder = load_rnn_models(sub_model_dir)
preds = predict_rnn.predict_samples(encoder, decoder, inputs_test,
output_len, n_classes)
else:
raise RuntimeError(f"Unsupport model type {model_type}")
events_num = pp.get_event_num(outputs_test, events_set_)
evals_summary, evals_separate, eval_events = evaluate_sub_model(
preds, outputs_test, events_set_, events_num)
if len(evals_summary) == 0: # 若evals_summary无元素，则表示评估的事件在评估日期范围内没有发生
return None, None
for event, eval_separate in zip(eval_events, evals_separate):
event, event_num, false_report, recall, false_alert, tier_precision, tier_recall, bleu_score = eval_separate
event_num = events_num[event]
pgsql.insert_model_test(event, event_num, false_report, recall,
false_alert, tier_precision, tier_recall,
bleu_score, detail_id)
score_summary, bleu_summary, tier_precision_summary, tier_recall_summary, fr_summary, rc_summary, fa_summary = \
evals_summary
scores.append([
score_summary, bleu_summary, tier_precision_summary,
tier_recall_summary, fr_summary, rc_summary, fa_summary, detail_id
])
return scores, events_num