DoP

Code of the CIKM 2025 paper "Decoder-only Pre-training Enhancement for Spatio-temporal Traffic Forecasting".

---

Raw Data

Raw data can be found at https://github.com/GestaltCogTeam/BasicTS/tree/master/datasets, and unzip them to datasets/raw_data.

Requirements.txt

numpy 1.22.3
torch 1.13.1
easy-torch 1.3.2
scipy 1.10.1

Data Processing

Run the following codes to conduct data processing, (replace X with corresponding number, and set history_seq_len as sequence length)

For examples,

python scripts/data_preparation/PEMS0X/generate_training_data.py --history_seq_len 12
python scripts/data_preparation/PEMS0X/generate_training_data.py --history_seq_len 2016

Run pre-training

Run the following code to conduct pre-training,

python run.py --cfg 'dop/pretrain_PEMS0X.py' --gpus=0

You can set up multiple gpus for parallel computing.

Run Forecasting

After pre-training, set the pre-trained checkpoint path in forecast_PEMS0X.py ('pre_trained_path'). Run the following code to conduct forecasting training processing,

python run.py --cfg 'dop/forecast_PEMS0X.py' --gpus=0

You can also set up multiple gpus for parallel computing.

Acknowledgement

Our code is developed based on BasicTS, STEP, STD-MAE, thanks for their excellent works!