Update
- The previous MXNet implementation of the paper now lives in
archive/mxnet/, unchanged. Everything else in this repository is a new PyTorch (Lightning) implementation.- The dataset (PEMS03/04/07/08, Parquet) is public on HuggingFace and ModelScope; the code downloads it automatically on first use.
- Prerequisites:
uvand a GPU driver with CUDA 12 support.# ModelScope (the default in every config) uv sync --extra ms uv run python scripts/train.py fit --config config/PEMS04/individual_GLU_mask_emb.yaml # HuggingFace uv sync --extra hf uv run python scripts/train.py fit --config config/PEMS04/individual_GLU_mask_emb.yaml \ --model.hub_source hf --data.hub_source hf
AAAI 2020. Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial-Temporal Network Data Forecasting
url: paper/AAAI2020-STSGCN.pdf
The original MXNet sources now live in archive/mxnet/; the code at the repository root is a PyTorch reimplementation that follows them exactly (same model, training loop and evaluator) and reproduces the paper's results.
Requires uv and a GPU driver with CUDA 12 support. The dataset is downloaded automatically on first use.
- install uv
- install the dependencies using
uv sync --extra ms - run code using
uv run python scripts/train.py fit --config config/PEMS04/individual_GLU_mask_emb.yaml
Use uv sync --extra hf and append --model.hub_source hf --data.hub_source hf to load the dataset from HuggingFace instead of ModelScope. The docker workflow of the MXNet version is kept in archive/mxnet/README.md.
| name | description |
|---|---|
| archive/mxnet | the original MXNet implementation (reference, unmodified) |
| compare | MXNet-vs-PyTorch forward parity harness |
| config | configurations of STSGCN |
| scripts | code of training STSGCN |
| stsgcn_torch | core of STSGCN |
| tests | pytest files |
| pyproject.toml | python packages requirements |