AngeliaFang/Traffic_Prediction

Traffic Prediction in PaddlePaddle (ASC17 Deep Learning Application)

Traffic Prediction

A Deep Learning Prediction Task Supported by PaddlePaddle

Introduction

1. Construct a easy Graph model by combine the adjacent nodes together

   • Collect spatio information from graph.csv, then construct a Center- MultiOrder-Adjacent Nodes Model (中心-多阶邻接节点图模型)
   • MultiOrder adjacent nodes can lengthen the predicting period

2. ResLSTM (Residual Long Short Term Memory)

   • Inpired by ResNet(which can deepen the neural network to more than 150 layers) (Deep Residual Network)
   • LSTM can deal with time sequence well
   • Residual can contribute to the loss and reasonable gradient

3. Construct the full network

   • Input speed sequences
   • Bidirectional LSTM can learn from the former speed and latter speed to deal with the relation of the adjacent nodes and the center node.
   • Use fully connected layers to extract the spatio feature.
   • End-to-End Trainable
   • LSTM and ReLU can lower the possibility of Gradient Vanishing

Usage

Install PaddlePaddle

go to the PaddlePaddle documentation

Prepare data

cd data

• get point list for training (Graph Constructing)

points_to_point()
get_points_count_list_2()

• remove zeros

split_data.split_by_remove_some_timestamps('speeds.csv',[(from_time, to_time)],'speed_nzero.csv')

• validation set

split_data.split_out(filename, [(from_time, to_time)],[output filenames])

• test data

split_data.get_test_data('test_speeds.csv', 'train_speeds.csv', 'speed_no_valid.csv', [(from_time, to_time)])                                                                                   dataset = create_dataset('test_speeds.csv')

• create data set

dataset = create_dataset('test_speeds.csv')
get_speed_data_2(dataset, 'test')

dataset = create_dataset('train_speeds.csv')
get_speed_data_2(dataset, 'train')

dataset = create_dataset('speeds_without_zero.csv')
get_predict_data_2(dataset)
dataset = create_dataset('VadiationSet/419_6_10.csv')
get_predict_valid(dataset)

run

sh run.sh

Advantages

• End-to-End trainable
• Maybe high accuracy :)
• Generalization ability and robustness
• make full use of data provided

Drawbacks

• training speed
• Framework (PaddlePaddle) is too slow and stupid

Licence

This project is under the Apache-2.0