davidham3 / stsgcn

AAAI 2020. Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial-Temporal Network Data Forecasting

STSGCN

AAAI 2020. Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial-Temporal Network Data Forecasting

url: paper/AAAI2020-STSGCN.pdf

Usage

Docker is recommended.

1. install docker
2. install nvidia-docker
3. build image using cd docker && docker build -t stsgcn/mxnet_1.41_cu100 .
4. download the data STSGCN_data.tar.gz with code: p72z
5. uncompress data file using tar -zxvf data.tar.gz
6. modify the term ctx in config/PEMS03/individual_GLU_mask_emb.json to match your GPU devices
7. run code using docker run -ti --rm --runtime=nvidia -v $PWD:/mxnet stsgcn/mxnet_1.41_cu100 python3 main.py --config config/PEMS03/individual_GLU_mask_emb.json

If you are using Microsoft OpenPAI, modify the configurations saved in the folder pai_jobs to train STSGCNs on your clusters.

repo structure

name	description
config	configurations of STSGCN
docker	dockerfile
models	core of STSGCN
pai_job	Microsoft OpenPAI configurations
paper	paper of STSGCN
test	pytest files
load_params.py	read parameters from local files
main.py	code of training STSGCN
pytest.ini	pytest configurations
requirements.txt	python packages requirements
utils.py	tools

stsgcn's Issues

Question about input_length in stsgcn()

I'm wondering what's the reason that you recursively set the input_length (time steps) in each STSGC layer to minus 2. Does that mean you set T=12 in the first layer, T=10 in the second layer, T=8 in the third layer and T=6 in the fourth layer? Is that for best predictive accuracy or for better hierarchical spatial-temporal logics? Thank you.
(more specific, line 487 in stsgcn.py)

about data

Hi~
thanks for your collected data, why is there 3 dimensions in pems08 and pems04 datasets, they are used to represent what?

Alternative data download?

Hi, I am unable to access baidu. Is your dataset hosted anywhere else? Thank you.

Num of vertices

Hello, sorry If I missed something in the paper but it seems that the current implementation is suitable with static graph (static number of nodes) right ?

about construct_adj

Thank you for your opening code. In utils.py construct_adj method, I found that you fill a same matrix in the diagnoal:
for i in range(steps): adj[i * N: (i + 1) * N, i * N: (i + 1) * N] = A
However, in your paper you said "he diagonal of the adjacency matrix are the adjacency matrices of the spatial networks of three continuous time steps".
Is there some other plcae you set the two continuous time steps metrix?

do you have the pytorch version？

We always use more pytorch, less mxnet. So i expect one version implemented by pytorch.

你的代码里，输出层这个线性是不是有错误

(B, N, C')

data = mx.sym.Activation(
    mx.sym.FullyConnected(
        data,
        flatten=False,
        num_hidden=num_of_filters
    ), 'relu'
)

这里的 num_of_filters 显然不是 C （特征）维度的长度？

Question about PEMS07

Sorry to disturb you, i have a question about the dataset PEMS07.

The time range of PEMS07 was 5/1/2017-8/31/2017 that lasted 123 days in your paper "AAAI20-STSGCN", but i found that there were only 98 days in the dataset you offered. Should the right time range of PEMS07 be 5/1/2017-8/6/2017?

Wish your reply soon, thanks.

Sensors' locations of PEMS, 03, 04, 07, 08

I'd appreciated if you could provide sensors' locations of PEMS, 03, 04, 07, 08

Weird testing results on METR-BAY and METR-LA

Hello,

I'd like to know if you tested once your model on METR-BAY and METR-LA?

From my side, your model shows incredibly bad results on the 2 datasets.

For instance, I got 12-step ahead forecasting results for METR-LA (the metric has been adjusted for loss error on 1 single value):
(MAE, MAPE, RMSE): (14.412544945659166, 34.03278628558422, 15.84779567776617)

Thanks in advance for your kindly response.

How to obtain the location of sensors

Only the relative distance between sensors is available from the csv file. How to obtain the location of sensors?

About node degree

Hi, I saw in your dataset, the average node degree is about one or two. What is the reason behind it?

environment

您好：
我安装了mxnet-cu101==1.5.1，在运行的时候总会出现下面的问题：

terminate called after throwing an instance of 'dmlc::Error'
terminate called recursively
what(): [10:39:57] /home/travis/build/dmlc/mxnet-distro/mxnet-build/3rdparty/mshadow/mshadow/./stream_gpu-inl.h:115: Check failed: err == CUBLAS_STATUS_SUCCESS (1 vs. 0) : Create cublas handle failed
Stack trace:
[bt] (0) /data/wangzeyu/wangzeyu_mxnet/lib/python3.7/site-packages/mxnet/libmxnet.so(+0x4b09db) [0x7f462d2089db]
[bt] (1) /data/wangzeyu/wangzeyu_mxnet/lib/python3.7/site-packages/mxnet/libmxnet.so(+0x25aa13b) [0x7f462f30213b]
[bt] (2) /data/wangzeyu/wangzeyu_mxnet/lib/python3.7/site-packages/mxnet/libmxnet.so(+0x25aaf78) [0x7f462f302f78]
[bt] (3) /data/wangzeyu/wangzeyu_mxnet/lib/python3.7/site-packages/mxnet/libmxnet.so(+0x25c5877) [0x7f462f31d877]
[bt] (4) /data/wangzeyu/wangzeyu_mxnet/lib/python3.7/site-packages/mxnet/libmxnet.so(+0x25c5aa6) [0x7f462f31daa6]
[bt] (5) /data/wangzeyu/wangzeyu_mxnet/lib/python3.7/site-packages/mxnet/libmxnet.so(+0x25c0a64) [0x7f462f318a64]
[bt] (6) /data/wangzeyu/wangzeyu_mxnet/bin/../lib/libstdc++.so.6(+0xc8421) [0x7f467ba11421]
[bt] (7) /lib/x86_64-linux-gnu/libpthread.so.0(+0x76ba) [0x7f46814a76ba]
[bt] (8) /lib/x86_64-linux-gnu/libc.so.6(clone+0x6d) [0x7f46811dd4dd]

已放弃 (核心已转储)
请问您之前遇到过这种问题吗？

Baidu netdisk issue

Is it possible that you to provide your datasets on google drive?

关于Astgcn(r)

请问ASTGCN(r) 的代码是将train_r输入ASTGCN_submodule后直接输出吗？

about result

come again~
I have run your data on my model, but the result is unbelievable. I almost get the same result with the model graph wavenet on public dataset metr-la. But on your dataset PEMS08, I get MAE: 1.57, MAPE: 3.44%, RMSE: 3.73. Your results are MAE 17.13, MAPE 10.96%, RMSE 26.80.
which is much smaller. I have checked my code, but I didn't find any problem. Do you know why? can you tell me your experiment details?

about constructing graph

How to construct a graph that conforms to the actual distribution with only detector coordinates?
Is there any written code, or can you explain the idea?
thank you ！

does everyone know in this paper, How many features are used in the experimental setting?

i saw one features in PEMS 03 but 08 has three features.