def main(argv):
t = time.time()
dnn_feature_columns, linear_feature_columns = get_feature_columns()
stage = argv[1]
print('Stage: %s'%stage)
eval_dict = {}
predict_dict = {}
predict_time_cost = {}
ids = None
for action in ACTION_LIST:
print("Action:", action)
model = WideAndDeep(linear_feature_columns, dnn_feature_columns, stage, action)
model.build_estimator()
if stage in ["online_train", "offline_train"]:
# 训练 并评估
model.train()
ids, logits, action_uauc = model.evaluate()
eval_dict[action] = action_uauc
if stage == "evaluate":
# 评估线下测试集结果，计算单个行为的uAUC值，并保存预测结果
ids, logits, action_uauc = model.evaluate()
eval_dict[action] = action_uauc
predict_dict[action] = logits
if stage == "submit":
# 预测线上测试集结果，保存预测结果
ids, logits, ts = model.predict()
predict_time_cost[action] = ts
predict_dict[action] = logits
if stage in ["evaluate", "offline_train", "online_train"]:
# 计算所有行为的加权uAUC
print(eval_dict)
weight_dict = {"read_comment": 4, "like": 3, "click_avatar": 2, "favorite": 1, "forward": 1,
"comment": 1, "follow": 1}
weight_auc = compute_weighted_score(eval_dict, weight_dict)
print("Weighted uAUC: ", weight_auc)
if stage in ["evaluate", "submit"]:
# 保存所有行为的预测结果，生成submit文件
actions = pd.DataFrame.from_dict(predict_dict)
print("Actions:", actions)
ids[["userid", "feedid"]] = ids[["userid", "feedid"]].astype(int)
res = pd.concat([ids, actions], sort=False, axis=1)
# 写文件
file_name = "submit_" + str(int(time.time())) + ".csv"
submit_file = os.path.join(FLAGS.root_path, stage, file_name)
print('Save to: %s'%submit_file)
res.to_csv(submit_file, index=False)
if stage == "submit":
print('不同目标行为2000条样本平均预测耗时（毫秒）：')
print(predict_time_cost)
print('单个目标行为2000条样本平均预测耗时（毫秒）：')
print(np.mean([v for v in predict_time_cost.values()]))
print('Time cost: %.2f s'%(time.time()-t))
def main(argv):
stage = argv[1]
eval_dict = {}
predict_dict = {}
ids = None
submit = pd.read_csv(ROOT_PATH + '/test_data.csv')[['userid', 'feedid']]
for action in ACTION_LIST:
print("-------------------Action-----------------:", action)
model = lgb_ctr(stage, action)
if stage == "offline_train":
# 离线训练并评估
ids, logits, action_uauc = model.train_test()
eval_dict[action] = action_uauc
predict_dict[action] = logits
elif stage == "online_train":
# 评估线下测试集结果，计算单个行为的uAUC值，并保存预测结果
ids, logits = model.train_test()
predict_dict[action] = logits
else:
print("stage must be in [online_train,offline_train]")
#
if stage == "offline_train":
print(eval_dict)
weight_dict = {
"read_comment": 4,
"like": 3,
"click_avatar": 2,
"favorite": 1,
"forward": 1,
"comment": 1,
"follow": 1
}
weight_auc = compute_weighted_score(eval_dict, weight_dict)
print("Weighted uAUC: ", weight_auc)
if stage == "online_train":
# 计算所有行为的加权uAUC
actions = pd.DataFrame.from_dict(predict_dict)
print("Actions:", actions)
ids[["userid", "feedid"]] = ids[["userid", "feedid"]].astype(int)
res = pd.concat([ids, actions], sort=False, axis=1)
# 写文件
submit_file = "./submit_lgb_6_6.csv"
print('Save to: %s' % submit_file)
res.to_csv(submit_file, index=False)
assert args.action in ACTION_LIST # read_comment,forward,like,click_avatar,comment,follow,favorite
score = {}
for action in ACTION_LIST:
print(f"------------------Training {action}------------")
score[action] = main(root_path=args.root_path,
stage=args.stage,
model_name=args.model_name,
action=action,
epoch=args.epoch,
learning_rate=args.learning_rate,
batch_size=args.batch_size,
weight_decay=args.weight_decay,
device=args.device,
save_dir=args.save_dir)
wauc = compute_weighted_score(score, WEIGHT_AUC)
print(f"weight uAUC:{wauc}")
with open('result.txt', 'a+') as f:
f.write(f'{args.model_name} wAUC {wauc}\n')
print(f'test wAUC: {wauc}')
# submit(
# root_path=args.root_path,
# stage="submit",
# model_name=args.model_name,
# action="all",
# epoch=args.epoch,
# learning_rate=args.learning_rate,
# batch_size=args.batch_size,
# weight_decay=args.weight_decay,
# device=args.device,