Recurrent Autoencoder with Skip Connections and Exogenous Variables for Traffic Forecasting

This is the implementation of the paper "Recurrent Autoencoder with Skip Connections and Exogenous Variables for Traffic Forecasting" submitted to the Traffic4cast Challenge 2019 at NeurIPS (https://www.iarai.ac.at/traffic4cast/).

Documentation

Our documentation is available in arxiv!

A full paper is now available in the Proceedings of Machine Learning Research. PMLR, 2020.

Summary

Given a sequence of frames of length q in {1,2, ..., 12} in a video, it can predict the next 3 consecutive frames.

Our model accepts a sequence of any size as input by iteratively using the encoder and concatenating its outputs. Then a recurrent encoder accumulates the temporal information of the input sequence into a single representation, and a recurrent decoder gives us three embedded predictions. Afterward, these predictions are upsampled to the original space by a decoder that uses skip connections from each layer of the encoder, but only from the last frame in the input sequence including the image. The Embedding Loss makes the recurrent layer GRU together with the encoder to produce better predictions in a low dimensional space. Frames Loss and skip connections from sibling layers in the encoder empowers the decoder to produce outputs with high definition. Exogenous variables are concatenated with the encoder output plus a fully connected layer before recurrent layers.

Training

• The notebook 'Train.ipynb' can be used to train any model except the one explained bellow, evaluate them, and even generate a submission for the challenge with the proper format.
• The notebook 'train-RAE_Clf' allows training the model called in the paper as 'RAE_Clf'
• The notebook 'exogenous_data.ipynb' shows figures of aggregated traffic. To use it, change the route in src/exogenous_data.py, variable DAYS_INFO_PATH to point agg_data/1_regions_features_mean.npy
• The folder src contains all the scripts that allow the notebooks to run:
  • src/data.py: Contains the class dataset that allows different sampling strategies. To use it, just change the variable DATA='/home/pherruzo/data/nips_traffic/' to the folder containing the data in your machine.
  • src/models/* : Our best model (image above) can be found in file 'RAEwSCwWSwIN.py', as well as 'RAE_Clf' in 'RAEwSCwWSwINxCLF.py'
  • models.py: Baseline models and basic autoencoder can be found here
  • exogenous_data.py: Contains all exogenous data retrieval
  • charts: Contains functions to plot aggregated traffic
  • losses.py: Contains loss functions definition and methods to compute the performance of the models

Cite

When referencing Recurrent Autoencoder with Skip Connections and Exogenous Variables for Traffic Forecasting, please cite this paper.

@InProceedings{pmlr-v123-herruzo20a, title = {Recurrent Autoencoder with Skip Connections and Exogenous Variables for Traffic Forecasting}, author = {Herruzo, Pedro and Larriba-Pey, Josep L.}, pages = {47--55}, year = {2020}, editor = {Hugo Jair Escalante and Raia Hadsell}, volume = {123}, series = {Proceedings of Machine Learning Research}, address = {Vancouver, CA}, month = {08--14 Dec}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v123/herruzo20a/herruzo20a.pdf}, url = {http://proceedings.mlr.press/v123/herruzo20a.html}, abstract = {The increasing complexity of mobility plus the growing population in cities, together with the importance of privacy when sharing data from vehicles or any device, makes traffic forecasting that uses data from infrastructure and citizens an open and challenging task. In this paper, we introduce a novel approach to deal with predictions of volume, speed and main traffic direction, in a new aggregated way of traffic data presented as videos. Our approach leverages the continuity in a sequence of frames, learning to embed them into a low dimensional space with an encoder and making predictions there using recurrent layers, ensuring good performance through an embedded loss, and then, recovering back spatial dimensions with a decoder using a second loss at a pixel level. Exogenous variables like weather, time and calendar are also added in the model. Furthermore, we introduce a novel sampling approach for sequences that ensures diversity when creating batches, running in parallel to the optimization process.} }

Contact

If you need help, links to the dataset, or scripts to download the exogenous data, do not hesitate to contact us: pherruzo@ac.upc.edu