zezhishao / stid Goto Github PK

Hi，dear author！Thanks for ur great work STID！Im a rookie in ST field，I want to deploy my own dataset with STID model，could u please tell me how to onfigure my own dataset？
My dataset time span from 2019-2020,every one hour for one record, count for 23347 nodes.
Thanks!

您好！看到您的论文，感觉大道之简，但本人不太熟悉时空预测领域问题，想请教一下：

为什么要将时间信息和空间信息作为embedding去学，而不直接将时间信息和空间信息作为特征输入呢？是因为该领域的数据集没有时间信息和空间信息么，如果是这样，方便问下该领域数据集的输入特征是什么吗？

您好，模型开始前的代码是这样的，备注是我自己写的维度：
batch_size, _, num_nodes, _ = input_data.shape # B,L,N,C
input_data = input_data.transpose(1, 2).contiguous() # B,N,L,C
input_data = input_data.view(batch_size, num_nodes, -1).transpose(1, 2).unsqueeze(-1) # B,N,(LC)->B,(LC),N->B,(LC),N,1
time_series_emb = self.time_series_emb_layer(input_data)
如果我没理解错的话L是窗口长度，C是每个节点的dim，请问这里为什么把L，C合并了，而不是把NC合并呢？谢谢

PEMS 系列数据集包括：PEMS03、PEMS04、PEMS07、PEMS08。为什么没有用PEMS03呢？谢谢。

I noticed that only the last time step of each sample is kept by the following code:

More specifically, the operation "t_i_d_data[:, -1, :]" is used to get the last time step from input data.

Could you please provide more interpretations?

请问作者代码中
self.scaler = load_pkl("{0}/scaler_in{1}_out{2}.pkl".format(cfg["TRAIN"]["DATA"]["DIR"], cfg["DATASET_INPUT_LEN"], cfg["DATASET_OUTPUT_LEN"]))
使用了这个文件在哪里可以下载

self.hidden_dim = self.embed_dim+self.node_dim * \
            int(self.if_spatial)+self.temp_dim_tid*int(self.if_day_in_week) + \
            self.temp_dim_diw*int(self.if_time_in_day)

维度相加的代码，if_day_in_week是和tid相乘吗？好像乘反了，会影响消融实验的？希望作者看下

为什么安装easy-torch=1.2.10后，安装的pytorch1.10.0被卸载了。

from basicts import launch_training
首先，上述会报错。basicts内部似乎没有launch_training
即使改成from basicts.launcher import launch_training，也无法运行额。
更为重要的是，不知道是如何落实到训练和预测的。感觉缺失了一部分，只有数据预处理和模型骨架和参数设置的部分。可能是我没有找到，还望指点。

Hi：
I tried to run your code, but I didn't run through
TypeError: forward() missing 4 required positional arguments: 'future_data', 'batch_seen', 'epoch', and 'train'
Thanks so much for your reply.

I noticed you take the last slice in the L dimension. In my understanding, T^TiD will change in L dimension. Why don't use t_i_d_data[:, :, -1]? temporal embeddings should have nothing to do with num_nodes

您好，有关于空间Identity，我只看到了随机生成的空间特征的矩阵，但是比如说PEMS04数据集，他的adj_mx，距离矩阵是不是没有用上？还是我在什么地方遗漏了。还望解答，谢谢

# prepare data
input_data = history_data[..., range(self.input_dim)]
...
# time series embedding
batch_size, _, num_nodes, _ = input_data.shape
input_data = input_data.transpose(1, 2).contiguous()
input_data = input_data.view(batch_size, num_nodes, -1).transpose(1, 2).unsqueeze(-1)
time_series_emb = self.time_series_emb_layer(input_data)

I want to know whether the shape of "input_data" in input_data = input_data.transpose(1, 2).contiguous() should be (B,N,L,1). But now it is (B,N,L,3). And i think the code input_data = history_data[..., range(self.input_dim)] should be replaced with input_data = history_data[..., 0].

Dear shao, in your paper,

For T DiW ∈ R𝑁𝑤 ×𝐷 , where 𝑁𝑤 = 7 ≪ 𝐷 = 32, we train STID by setting the embedding size of T DiW to 2 to get a more accurate visualization.

And in your code, # concate all embeddings hidden = torch.cat([time_series_emb] + node_emb + tem_emb, dim=1)

My confusion is how to concatenate the three embeddings in the given dimension when training the model if you change the embedding size of T DiW to 2 whileas the size of other embeddings is still 32.
Other embeddings size change to 2 also? Or something else?

node_emb仅仅对随机生成的矩阵进行交互，但是不直接对his_data进行embed。虽然从知乎上看到这个是一直随机embed方式，但是还是从直觉上很难接受，his_data的空间效应没有直接被编码。如果采用nn.Embedding对空间信息同时进行embed效果会怎么样呢，请问这个方面有对比过吗？或者说，直接embed空间效应有什么编程上的困难吗？

Great project!
However, when I try to get a prediction with shape [B, L, N, C] through STID, I can only get shape [B, L, N, 1].
During the time series embedding process, the input dim C is reduced and unsqueezed with 1. How to keep the output dim the same as C.