Baidu Traffic speed prediction

This repo is implementation of "Baidu Traffic Speed Prediction"

A random mechanism is used in this implementation, program will choose Model from self-attention, GRU, 1-5 hidden layers Dense net, also random choose hyper parameters and random choose data features to train.

Further idea is using nerual arhitecture search to do model evolution.

Requirement

• python 3.4+
• pytorch 0.4.1+
• tqdm
• numpy

Usage

0) Download the data.

Download data in this link

In this notebook only use traffic_speed_sub-dataset.zip and road_network_sub-dataset.zip packages.

1) Preprocess the data.

Create data folder, and unzip upper two packages to it

mkdir data
unzip traffic_speed_sub-dataset.zip -d ./data
unzip road_network_sub-dataset.zip -d ./data

create clean folder

mkdir clean

run data processing scripts

python 01_trafficDataRaw.py
python 02_extract_link_id_dict.py
python 03_geoInfo.py
python 04_timeInformation.py
python 05_perpare_train_data.py

dataset is splited into first half and second half, this notebook will only use first half to train and validation.

2) Randomly choose models and features and train

Example:

python autoML.py SimpleFC 1      # choose SimpleFC model and try once.
python autoML.py SimpleFC3 5     # choose SimpleFC3 model and try 5 times.
python autoML.py Transformer 1    # choose Transfomer model and try once.
python autoML.py GRU 1           # choose GRU model and try once.

currently support model including: SimpleFC(1 hidden layer Dense net), SimpleFC2(2 hidden layer Dense net) ... to SimpleFC5(5 hidden layer Dense net), SimpleFC5_block(residential 9 hidden layer Dense net.), Transformer(multi head self-attention), GRU.

3) View result

step 2 will generate one folder for each try, run plot_result.ipynb to view results.