Temporal Fusion Transformer using PaddlePaddle

Based on Paddle implementation paper: Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting

datasets：Electricity dataset

You can run it on aistudio

You can download the best model we have trained from Baidu Pan

• link：https://pan.baidu.com/s/1XpOY04lt37NlzcEMM3XZEQ
• code：zjvu

The training steps:

• Convert TXT files to CSV files

• We have already provided a csv file dataset on aistudio so you can skip this step

  python scripts/convert_data.py
  

• Verifying configuration files

  python config/conf.py
  

• Verifying dataset

  python dataset/ts_dataset.py
  

• Training with single GPU

  python main.py --exp_name electricity --conf_file_path your_file_path --inference False
  

• Inference with best model saved

  python main.py --exp_name electricity --conf_file_path your_file_path --inference True
  

The experimental results

The Retrieval results

	paper	tensorflow（source code）	paddlepaddle
P50	0.055	0.056	0.054
P90	0.027	0.030	0.028

The reference material

• google-research/tft
• stevinc/Transformer_Timeseries