-
Notifications
You must be signed in to change notification settings - Fork 15
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
复现问题 #9
Comments
针对复现问题:从你测试结果来看,还没比过HA的性能,可能问题有以下几点:
使用新数据:网格划分大小没有严格要求,一个城市大概有几百个网格应该都是可以的。如果部分网格无数据,可以用mask的方式,attention计算不关注mask网格,计算loss不考虑mask网格。 |
感谢!第一个问题我把FewRatio改为1再重新训练下。 |
还有一个问题:Periodical Data的生成可以举个例子么,看README里面介绍的是有一个维度是Number of past days corresponding to the time of day,然后看了下Json数据里面这个数都是3,意思是取当前时间过去3天相同时刻的数据么? |
@kakaroy123 Hi, did you eventually manage to replicate the results? |
嗯 基本复现了 感谢! |
Could you please share some insights about what you did to make it work? Because I also got some results that are worse than HA. |
作者您好!关于无效数据加mask的方式还想请教一下,看了下代码,在Embedding后数据维度和数值都发生变化,无效数据的位置好像找不到在哪里了,怎么加mask呢? |
您好,无效数据是指什么意思? |
网格里一些无法获取真实数据而用0填充的数据。 |
做了下复现,还有几个问题想请教一下:
1、Crowd测试结果不一致问题
预训练指令:
python main.py --device_id 3 --machine machine --dataset CrowdCellularBikeNYCTaxiNYCBikeNYC2TaxiNYC2TrafficCDTrafficHZTrafficJNTrafficNJTrafficSHTrafficTJTrafficZZ --task short --size middle --mask_strategy_random 'batch' --lr 3e-4 --used_data 'diverse' --prompt_ST 0
得到预训练模型后,用Crowd数据集做了测试,指令如下:
python main.py --device_id 2 --machine machine --task short --size middle --prompt_ST 1 --pred_len 6 --his_len 6 --num_memory_spatial 512 --num_memory_temporal 512 --prompt_content 's_p_c' --dataset Crowd --used_data 'diverse' --file_load_path Pretrain_Dataset_CrowdCellularBikeNYCTaxiNYCBikeNYC2TaxiNYC2TrafficCDTrafficHZTrafficJNTrafficNJTrafficSHTrafficTJTrafficZZ_Task_short_FewRatio_0.5 --few_ratio 0.0
测试结果如下:
stage:0, epoch:0, best rmse: 0.002725724898491081
{'Crowd': {'temporal': {0.5: {'rmse': 15481.58774870661, 'mae': 7934.287060662857}}}}
我看论文里面,Crowd的短时预测,RMES是3000,MAE是1380,感觉和我测试结果还有较大差异,想问下可能是哪里的问题呢?
2、如果想拿些新的数据来做测试或者是预训练,网格划分大小有要求么?如果部分网格内的数据获取不到怎么处理?
The text was updated successfully, but these errors were encountered: