wengwenchao123/ddgcrn

[Pattern Recognition] Decomposition Dynamic Graph Conolutional Recurrent Network for Traffic Forecasting

ddgcrn's Issues

请问，在数据集PEMS03上得到的MAPE值非常大，是怎么回事呢？

Apologies for the delayed response! I was able to resolve this issue with your assistance :)"

          Apologies for the delayed response! I was able to resolve this issue with your assistance :)"

Originally posted by @seunghan96 in #3 (comment)

Missing argument "days_per_week" in Args

I have fixed the issue myself, but I wanted to remind that "days_per_week" is missing in line 103 of the file "lib/dataloader.py".
You may need to add the following in the "run.py" file so that the code can be run swiftly.
args.add_argument('--days_per_week', default=config['data']['days_per_week'], type=int)

Unable to obtain the desired pretrained weight

Hi :)
When running the code according to the author's instructions, the results from the table are not reproducible.
I would like to inquire if there is an issue with the code that might be causing this inconsistency.

Confusion about the code according to "Model Training" part

I'm confused about the code under the section "Model Training". I wonder if the right code should be "python run.py --dataset {DATASET_NAME} --mode {MODE_NAME}" instead of "python run.py --datasets {DATASET_NAME} --mode {MODE_NAME}". When using the code you provided, an error of "run.py: error: unrecognized arguments: --datasets PEMSD4" would occur.
I look forward to your reply.

Using the approach on new traffic counter datasets

Hi @wengwenchao123, I'm wondering about using the approach to estimate traffic on roads based on this input: https://roadtraffic.dft.gov.uk/

I've had a look at the code but not sure where to start, do you have an example of the structure that the input dataset needs to be in before running predictive model?

Thanks for sharing.

Dataset normalization issue

Thanks for your good work. In the code, the raw data is normalized by the mean and std values calculated from the whole dataset. This can cause an information leakage from the testing part and may be unfair for other baselines. A better choice could be calculating the mean and std using only the training and validation parts.

Question about hyper-parameter settings

Thanks for your great work and the sharing of the code. I have been recently running the script but I found that some hyper-parameter settings are inconsistent with those specified in the paper. For instance, the epoch and learning rate for PEMSD8 is 300 and 0.003 respectively given in the configuration file, while they are specified as 100 and 0.03 in the paper. May I ask which one should we follow? Thank you.

Dataset

您好，请问本项目使用的数据集中“PEMS04.csv”文件中的“from,to,cost”分别代表什么意思呢？

想请教一个问题

AGCRN里反归一化的我看是label，这里反归一化是output 我不太明白区别在哪 或者是我代码看的不够仔细 希望得到您的帮助 您说后面两年发的顶会指的是哪一篇呢

Segmented learning training strategy

您好，请问您使用分段学习了吗，我看到源码的第一层训练并没有执行？望解答