Dataset | Model | MAE | Parameter |
---|---|---|---|
METR-LA | STGCN(pytorch) | 3.982 | Epoch = 1000 |
Dataset | Node num | Time | Duration | Time slot | Scene |
---|---|---|---|---|---|
METR-LA | 207 | 2012.03.01~2012.6.27 | 4 months | 5mins | Loop detecors in highway |
PEMS-BAY | 325 | 2017.01.01~2017.06.30 | 6 months | 5mins | Sensors in Bay Area |
PEMSD7 | 228 | Workday of 2012.05-2012.06 | 44 days | 5mins | Sensors in California |
- METR-LA
Data | Info |
---|---|
distance_la_2012.csv | 两两节点之间的距离 |
graph_sensor_ids | 所有节点的id list |
graph_sensor_locations | 所有节点的坐标 |
metra_la.csv | 每个节点在每个时刻的速度信息 |
Q_Traffic Dataset Link
The data provider gives 15073 central road and its neighbour information, so there are totally 45148 roads data(speed/road netwok/gps) provided. The total time slot number is 5856(61days * 24hours * 4quarter).
Filename | Dimension | Instance | Tips |
---|---|---|---|
traffic_speed_sub-dataset | 3 * (5856*45148) | road_id = 1562548955, timeslot_id = 0, speed = 41.3480687196 | No headings, sep = ' ' |
road_network_sub-dataset | 8 * 45148(-Heading) | road_id = 1562548955, width = 30, direction = 3, snodeid = 1520445066, enodeid = 1549742690, length = 0.038, speedclass = 6, lanenum = 1 | Headings, sep = '\t' |
link_gps | 3 * 45418 | road_id = 1562548955, longtitude = 116.367557, latitude = 39.899537 | No headings, sep = ' ' |
query_sub-dataset | 61 * 6 * N | search_time = 2017-04-01 19:42:23, start_pos = (116.325461 40.036083), end_pos = (116.350811 40.090999), travel_time = 33 | No headings, sep = ' ' or ',' |
neighbours_1km.txt | 15073 * 11 | road_id = xx, pre1, pre2, ..., pre5, next1, next2, ..., next5 |
Highways England network journey time and traffic flow data Link
- Data pre-processing
- speed_h5py.py is used to generate speed dataset in h5 format.
- Then, use this speed_dataset to generate train/validate/test data by code generate_training_data.py .
- gen_adj_mx.py is used to generate road_map.
- Train DCRNN model
- Comand line(The version of tensorflow-gpu must be higher than tensorflow):
tmux a -t dcrnn_baidu
source activate python3.6
cd ~/workspace/GCN/DCRNN-master
python dcrnn_train.py --config_filename=data/model/dcrnn_baidu.yaml
- ChebNet: Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
- STGCN: Spatio-Temporal Graph Convolutional Networks | For pytorch version: pytorch version
- DCRNN: Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting
- Multi-head Self Attention Model(AutoInt): AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks CSDN reference: AutoInt:使用Multi-head Self-Attention进行自动特征学习的CTR模型
- K-SVD in Dictionary learning There are codes and some illustration.
- [osmnx guide]https://github.com/gboeing/osmnx-examples/tree/master/notebooks()
- python GIS
version | Python version | cuDNN | CUDA |
---|---|---|---|
tensorflow-gpu-1.14.0 | python3.6 | 7.6 | 10.0 |