Jittor Warmup1 Cora GCN

This repository contains the open-source code for the Jittor competition warmup task: node classification on Cora with a two-layer GCN and optional label propagation post-processing.

Environment

Recommended environment:

• Python 3.8+
• Jittor
• JittorGeometric
• NumPy

Install dependencies:

pip install -r requirements.txt

If JittorGeometric requires extra system or CUDA configuration, follow the official installation guide: https://github.com/AlgRUC/JittorGeometric#installation

Data

Download the official competition dataset and put it at:

data/cora.pkl

The repository does not include the raw dataset file. You can also place the dataset elsewhere and pass the directory with --data-root:

python -m src.cora_gcn --data-root /path/to/data

Expected dataset fields are documented in data/README.md.

Training

Run the default training and prediction pipeline:

python -m src.cora_gcn --config configs/default.json

Equivalent helper script:

bash scripts/train_and_predict.sh

The default run writes:

• outputs/default/config.json: actual config used by the run
• outputs/default/command.txt: command used to launch the run
• outputs/default/train.log: training and validation log
• outputs/default/result.json: prediction file for test nodes

Common overrides:

python -m src.cora_gcn \
  --config configs/default.json \
  --data-root data \
  --output-dir outputs/seed42 \
  --seed 42 \
  --hidden-dim 128 \
  --dropout 0.6

Command-line arguments have higher priority than config values.

Evaluation / Inference

The warmup task does not require a separate checkpoint-based inference step. The command above trains the model, selects the best validation epoch, applies label propagation when it improves validation accuracy, and writes result.json.

To package the result:

python scripts/package_submission.py \
  --result-path outputs/default/result.json \
  --zip-path outputs/result.zip

Result

Metric: node classification accuracy on the validation split during local development. The final online score may differ because labels for the test split are hidden by the competition platform.

Default model:

• 2-layer GCN
• hidden dimension: 128
• dropout: 0.6
• Adam learning rate: 0.01
• weight decay: 0.001
• early stopping by validation accuracy
• label propagation post-processing selected by validation accuracy

Reproducibility

Use --seed or the seed field in configs/default.json. The script sets Python, NumPy, and Jittor seeds before loading data and training.

If files are missing, the script reports the expected path and suggests using --data-root.

Third-party Dependencies

This project uses Jittor and JittorGeometric APIs, including GCNConv, gcn_norm, cootocsc, and cootocsr. Please follow their respective licenses when redistributing this code or building upon it.