This is the neural network model we at SPARTA built to infer vehicle speed on a certain link in the Xi'an road network given Didi data in the local area (sans that link) as per the TRANSFOR19 competition. See full competition details and results here and refer to both the Jupyter notebooks and SPARTA_transfor.pdf for further details on our model.
Some info and details that might be helpful for us.
git pull/push origin masterto push to and pull from our own Github repo- To grab updates from the competition board, do:
git remote add og https://github.com/TRANSFORABJ70/TRANSFOR19.gitto add a reference to the official repo (only need to do this once)git pull og masterto actually grab the changes and merge them into our repo
Generally we want to avoid changing any files that the official repo controls (e.g. area.png, Predictions.zip, and README.md).
The original datafile for December 1, 2016 whose average speeds we were asked to predict is too big to commit so here it is. Drop it into the main directory and git will be preset to ignore it.
From the board's email introducing the data:
In Github, we have posted the Predictions file containing 2 csv files (North and South direction). Predictions should be generated for the time periods 6:00 to 11:00 and 16:00 to 21:00 (marked with “x”). You can also find the entire city trajectories for the same day (gps_20161201) here.
We can request the full dataset for October and November from the Didi's Gaia Initiative webpage. These raw files go in data/tar/ and will be unzipped to /data/xian by 2a_data_download.ipynb.
If we use map information, make sure to use the eviltransform library to convert back and forth between standard projection coordinates (WGS-84) and the Chinese-specific projection coordinate system (GCJ-02).