T-GCN-PyTorch

T-GCN-PyTorch

This is a PyTorch implementation of T-GCN in the following paper: T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction.

A stable version of this repository can be found at the official repository.

Note that the original implementation is in TensorFlow, which performs a tiny bit better than this implementation for now.

Requirements

• numpy
• pandas
• torch
• lightning>=2.0
• torchmetrics>=0.11

⚠️ The repository is currently based on Lightning 2.0. To use PyTorch Lightning v1.x, please switch to the pl_v1 branch.

Model Training

• CLI

  # GCN
  python main.py fit --trainer.max_epochs 3000 --trainer.accelerator cuda --trainer.devices 1 --data.dataset_name losloop --data.batch_size 64 --data.seq_len 12 --data.pre_len 3 --model.model.class_path models.GCN --model.learning_rate 0.001 --model.weight_decay 0 --model.loss mse --model.model.init_args.hidden_dim 100
  # GRU
  python main.py fit --trainer.max_epochs 3000 --trainer.accelerator cuda --trainer.devices 1 --data.dataset_name losloop --data.batch_size 64 --data.seq_len 12 --data.pre_len 3 --model.model.class_path models.GRU --model.learning_rate 0.001 --model.weight_decay 1.5e-3 --model.loss mse --model.model.init_args.hidden_dim 100
  # T-GCN
  python main.py fit --trainer.max_epochs 1500 --trainer.accelerator cuda --trainer.devices 1 --data.dataset_name losloop --data.batch_size 32 --data.seq_len 12 --data.pre_len 3 --model.model.class_path models.TGCN --model.learning_rate 0.001 --model.weight_decay 0 --model.loss mse_with_regularizer --model.model.init_args.hidden_dim 64
  

• YAML config file

  # GCN
  python main.py fit --config configs/gcn.yaml
  # GRU
  python main.py fit --config configs/gru.yaml
  # T-GCN
  python main.py fit --config configs/tgcn.yaml
  

Please refer to python main.py fit -h for more CLI arguments.

Run tensorboard --logdir ./lightning_logs to monitor the training progress and view the prediction results.