My experiment notes

生成实验数据

• generate_training_data.py，代码会将数据格式变成(B, T, V, F)的四维tensor，放到final data下面
  • n_his，用多少个过去的点预测
  • n_pred，预测未来多少个点

跑实验

• train.py，实验结果，包括log和图像，都在result下面

Graph WaveNet for Deep Spatial-Temporal Graph Modeling

This is the original pytorch implementation of Graph WaveNet in the following paper: [Graph WaveNet for Deep Spatial-Temporal Graph Modeling, IJCAI 2019] (https://arxiv.org/abs/1906.00121).

Requirements

• python 3
• pytorch
• scipy
• numpy
• pandas
• pyaml

Data Preparation

Step1: Download METR-LA and PEMS-BAY data from Google Drive or Baidu Yun links provided by DCRNN.

Step2: Follow DCRNN's scripts to preprocess data.

# Create data directories
mkdir -p data/{METR-LA,PEMS-BAY}

# METR-LA
python -m scripts.generate_training_data --output_dir=data/METR-LA --traffic_df_filename=data/metr-la.h5

# PEMS-BAY
python -m scripts.generate_training_data --output_dir=data/PEMS-BAY --traffic_df_filename=data/pems-bay.h5

Experiments

Train models configured in Table 3 of the paper.

ep=100
dv=cuda:0
mkdir experiment
mkdir experiment/metr

#identity
expid=1
python train.py --device $dv --gcn_bool --adjtype identity  --epoch $ep --expid $expid  --save ./experiment/metr/metr > ./experiment/metr/train-$expid.log
rm ./experiment/metr/metr_epoch*

#forward-only
expid=2
python train.py --device $dv --gcn_bool --adjtype transition --epoch $ep --expid $expid  --save ./experiment/metr/metr > ./experiment/metr/train-$expid.log
rm ./experiment/metr/metr_epoch*

#adaptive-only
expid=3
python train.py --device $dv --gcn_bool --adjtype transition --aptonly  --addaptadj --randomadj --epoch $ep --expid $expid  --save ./experiment/metr/metr > ./experiment/metr/train-$expid.log
rm ./experiment/metr/metr_epoch*

#forward-backward
expid=4
python train.py --device $dv --gcn_bool --adjtype doubletransition  --epoch $ep --expid $expid  --save ./experiment/metr/metr > ./experiment/metr/train-$expid.log
rm ./experiment/metr/metr_epoch*

#forward-backward-adaptive
expid=5
python train.py --device $dv --gcn_bool --adjtype doubletransition --addaptadj  --randomadj  --epoch $ep --expid $expid  --save ./experiment/metr/metr > ./experiment/metr/train-$expid.log
rm ./experiment/metr/metr_epoch*