Various methods for sequence-to-sequence modeling have been reported in the previous blog post. While the results were promising, such models do not apply to a more complex traffic network for the following reason. Spatial dependence of traffic volume on an extended road plays an essential role in traffic flow. For example, it is known that the traffic flow at upstream routes impacts traffic status at downstream roads through the transfer effect. The same is valid in the opposite direction. Moreover, the interconnectedness of the network means that a traffic jam will have an impact with a radius around it.
Temporal Graphical Convolutional Networks (T-GCN) is great for problems such as time-series dataset on a complex network. A brief overview as to why T-GCN is a good candidate for traffic modeling is
- The “convolutional” network of the network captures trends in the traffic flow
- The “graphical” convolution takes into account the connectivity of the links
- The “temporal” part mitigates the non-causal issue that exists in the “convolutional” part.
As we see, different networks in the first part of the project are used in this architecture.
The main task for the last part was to port and successfully run the T-GCN model developed by Zhao et al.  on the Berzelius cluster. The TensorFlow implementation of T-GCN as is available in the GitHub repository  has been ported and tested on Berzelius. The version of the code fetched at the time of writing this article is TensorFlow 1.x, an older version of TensorFlow not readily available on Berzelius. Many Singularity containers with a correct version of TensorFlow and required software stacks have been created to deploy the code on the cluster.
Afterward, the code was successfully run and tested against the benchmark dataset readily available in the repository. Therefore, we conclude the porting of T-GCN to Berzelius (or any other cluster with Singularity runtime).
 L. Zhao, Y. Song, C. Zhang. Y. Liu, P. Wang, T. Lin, M. Deng, H. Li, "T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction," in IEEE Transactions on Intelligent Transportation Systems, 21, 3848-3858, (2020) doi: 10.1109/TITS.2019.2935152.
 GitHub: https://github.com/lehaifeng/T-GCN