Thank you for your awesome work!
Download from baidu is almost impossible.
Could you share dataset on another platform or e-mail?

thanks.

您好！感谢您的工作，我正在尝试您的代码。请问AttributeError: 'list' object has no attribute 'asnumpy'的问题该怎么解决？我用的是mxnet-cu100==1.7.0。请问您当初用的是哪个版本的？mxnet的版本不同似乎会造成非常多的问题。
Traceback (most recent call last):
File "main_4n0_3layer_12T_res.py", line 27, in
net = construct_model(config)
File "/work/users/s/h/shuaishu/Transformer/STFGNN/utils_4n0_3layer_12T_res.py", line 41, in construct_model
init=mx.init.Constant(value=adj_mx.tolist()))
File "/work/users/s/h/shuaishu/.conda/STFGNNpy37/lib/python3.7/site-packages/mxnet/symbol/symbol.py", line 2864, in var
init = init.dumps()
File "/work/users/s/h/shuaishu/.conda/STFGNNpy37/lib/python3.7/site-packages/mxnet/initializer.py", line 477, in dumps
self._kwargs['value'] = val.tolist() if isinstance(val, np.ndarray) else val.asnumpy().tolist()
AttributeError: 'list' object has no attribute 'asnumpy'

Figure 3(b) is the example of Spatial-Temporal Fusion Graph.这一段，有不止一个表述上的问题。

Hi, Thanks for your wonderful works~
Recently, I found there seems to be some inconsistency between STFGNN and other baselines (e.g., Graph WaveNet, DCRNN) in the way of calculating metrics, which may make the comparison unfair.
Specifically, most of the baselines calculate the metrics at horizons 3, 6, and 12, while STFGNN seems to calculate the metrics among horizons 1~3, 1~6, and 1~12.
Do I understand the code correctly?

作者，你好。我在PEMS03数据集上运行你们提供的demo，为什么运行出的结果在前2轮看起来很正常，到第3轮的train loss直接变成200多了，直到训练结束也还是200多，没有得到正确的实验结果。dtw文件就是adj_PEMS03_001.csv吧，我也尝试用fast_DTW_gen.py重新生成adj_PEMS03_001.csv，但还是在第3轮遇到同样的问题，请问原因是什么？

您好，请问您可以帮我解释一下这些参数分别代表什么嘛？还有程序的运行流程是怎么样呢？需要我自己在单独的去生成每个数据集时态图吗？

I got the following error while running the code on PEMS08. I guess the reason could be wrong version of my mxnet or python. Could you tell me your versions of mxnet and python?
The shape of localized adjacency matrix: (680, 680) Traceback (most recent call last): File "main_4n0_3layer_12T_res.py", line 27, in <module> net = construct_model(config) File "/home/wx/paper/20201119_/STFGNN-master/utils_4n0_3layer_12T_res.py", line 41, in construct_model init=mx.init.Constant(value=adj_mx.tolist())) File "/home/wx/miniconda3/envs/gcn/lib/python3.7/site-packages/mxnet/symbol/symbol.py", line 2864, in var init = init.dumps() File "/home/wx/miniconda3/envs/gcn/lib/python3.7/site-packages/mxnet/initializer.py", line 477, in dumps self._kwargs['value'] = val.tolist() if isinstance(val, np.ndarray) else val.asnumpy().tolist() AttributeError: 'list' object has no attribute 'asnumpy'

return mx.sym.max(mx.sym.concat(*need_concat, dim=0), axis=0)
I don't understand what is the meaning of *need_concat? What's the purpose of *?

大佬，想引用你的文章，请问你文章具体bibtex是什么呢？谷歌学术上只能搜到arxiv， 搜不到被AAAI录用的详情

I am trying to use Temporal_Graph_gen.py to process METR-LA.

README said that in PEMS08 dataset, the shape is (Total_Time_Steps, Node_Number), however, when I load the PEMS08, the shape is (17856,170,3), Could you please tell me what does the third dimension means?

你好，文中描述The set of α which determines how many smallest numbers are treated in Alg. 1 is tricky and we would analysis it in the section of experiments.但是再算法1的伪代码中没有找到α ，还想请问下大佬算法1中的第6行到第7行的代码表示什么意思

您好：
请问metr-la.csv文件如何获取？

Hello,

i am interesting in the prediction in quite large networks but there is not a lot of models for this purpose. I think your work is great state-of-the-art method and I would like to test my model on PEMS07 dataset, but there is no way to download it from Baidu.

Can you please send it to me? My emal address is [email protected].

Thank you

When will the code be released？

There is a lot of missing data (those with value 0) in several datasets (such as METR-LA). Did you adopt any interplolation approach (such as linear interplolation) to fill in these missing values?

从main.py第159行的y, x = test_y[:, : idx + 1, :], prediction[:, : idx + 1, :]可以看出使用的是预测时间序列平均值作为评价标准

之前的工作都是使用的预测序列的最后一个时间片的值作为评价标准

希望作者能公正比较，将论文中的结果改为最后一个时间片的值而不是平均值

请问你是如何复现出Graph wavenet的结果，我自己跑出Graph wavenet的结果比你论文中展示的小很多。

Hi, I'm trying to set the environment in colab.
I cloned this repo & installed packages with requirements.txt
After that, I ran below command as it's mentioned in the README.md.

python main_4n0_3layer_12T_res.py --config config/PEMS04/individual_3layer_12T.json

however, I'm getting AttributeError, which seems to be related with mxnet version issue

Traceback (most recent call last):
File "main_4n0_3layer_12T_res.py", line 27, in
net = construct_model(config)
File "/content/STFGNN/utils_4n0_3layer_12T_res.py", line 41, in construct_model
init=mx.init.Constant(value=adj_mx.tolist()))
File "/usr/local/lib/python3.7/dist-packages/mxnet/symbol/symbol.py", line 2970, in var
init = init.dumps()
File "/usr/local/lib/python3.7/dist-packages/mxnet/initializer.py", line 477, in dumps
self._kwargs['value'] = val.tolist() if isinstance(val, np.ndarray) else val.asnumpy().tolist()
AttributeError: 'list' object has no attribute 'asnumpy'

I think it may be related with mxnet version issue.
Could you clarify specific version information of each required packge you used in the project?
Or do you have any other possible cause for that Error?

Thank you in advance :)

Excuse me, could you please talk about the concatenation of the three adjacency matrices?

Excuse me Sir ,the process was stopped at here and I do not know the reason .Could you help me?

The shape of localized adjacency matrix: (1432, 1432)
(15701, 12, 358, 1) (15701, 12, 358)
(5219, 12, 358, 1) (5219, 12, 358)
(5219, 12, 358, 1) (5219, 12, 358)

Hi,thanks for your excellent work!
I have a question, is the performance of Gated CNN berter than TCN?

您好，由于您宝贵的框架基于STSGCN，所以我想问下关于STSGCN的复现问题。因为现在貌似大家都无法复现它，所以我想问下当时您的复现情况，是不是我们忽略了它的一些细节。若您能提供一些方向也可以，因为那个框架的复现结果和论文结果差的有点远，若您能提供您复现的方向把差距缩小到误差范围内，我们将感激不尽，谢谢。

作者，你好。我看你们code里是让T=12，然后分别求出前1个horizon的MAE，前2个horizon的MAE，...，前12个horizon的MAE，用前12个horizon（也就是整个label和预测值pred）的MAE，MAPE和RMSE作为论文的最终结果，我的理解没错吧？

In def stsgcn(data, adj, label, input_length, num_of_vertices, num_of_features, filter_list, module_type, activation, use_mask=True, mask_init_value=None, temporal_emb=True, spatial_emb=True, prefix="", rho=1, predict_length=12), mask_init_value is None.
However, if mask_init_value is None: is written in line 418 and raise ValueError("mask init value is None!") is written in line 419.
Isn't that contradictory?

大佬，您好，文章中A_{TC}的含义没有看懂， 为啥i=3N+l, and j=2N+l, A_{STFG(i,j)}=1呢

Hi,

Can you tell me the new results of STFGNN on METR-LA and PeMS-BAY datasets? I'd like to cite your paper, and I should report the performance of STFGNN on these two datasets. If you tell me the results, then I don't need to run your model myself.

Thank you very much!

你好，为了方便叙述和解释，我这里用中文说一下：

这里提到，每一个STFGN Module的输入都是一个原输入的切片，每个切片的长度为K，然后经过Maxpool之后，K变为1，那么每个Module的输出维度不应该是1Nd*C吗？为什么这里提到输出第一维是(T-K+1)？

How to solve it?

Traceback (most recent call last):
File "main_4n0_3layer_12T_res.py", line 27, in
net = construct_model(config)
File "/mnt/d/try/STFGNN-master/utils_4n0_3layer_12T_res.py", line 41, in construct_model
adj = mx.sym.Variable('adj', shape=adj_mx.shape,
File "/home/lyl/miniconda3/lib/python3.8/site-packages/mxnet/symbol/symbol.py", line 2970, in var
init = init.dumps()
File "/home/lyl/miniconda3/lib/python3.8/site-packages/mxnet/initializer.py", line 477, in dumps
self._kwargs['value'] = val.tolist() if isinstance(val, np.ndarray) else val.asnumpy().tolist()
AttributeError: 'list' object has no attribute 'asnumpy'

Thanks.

您好，mxnet 版本着实看不懂，有无pytorch版本呢

Hi,

I found that the Graph Wavenet results reported in your paper seem to be incorrect. I understand that you just use the results from STSGCN, but it would be better to reproduce the results by yourself.

From what I have reproduced, Graph Wavenet achieves an average MAE (average of 12 steps) of 14.8 on PEMS-03, an average MAE of 19.4 on PEMS-04, and an average MAE of 14.9 on PEMS-08.

想请问一下这个公式里各个元素的shape，不能理解。在STSGCN中有个一模一样的公式，但是那个shape很清楚。主要就是h的shape是多少不太明白。感谢