uyen1109/elliptic-fraud-deploy

Elliptic Fraud — HF Model Repo (Graph-ready)

This repo supports two payload schemas:

(A) Node mode (single prediction)

{
  "features": {
    "tx_value": 0.23,
    "in_degree": 4,
    "out_degree": 1,
    "age_days": 12.0
  }
}

• Uses feature_config.json to fix the key order (recommended). If not set, keys are sorted alphabetically.

(B) Graph mode (batch prediction over nodes)

{
  "x": [[0.23,4,1,12.0], [0.5,2,3,30.0], ...],      // [N,F]
  "edge_index": [[0, 1, 2, ...], [1, 2, 0, ...]]     // [2,E], row indices into x
}

• If MODEL_TYPE=graphsage and torch_geometric is available, uses GNN; otherwise falls back to MLP over x.

Environment variables on Endpoint

• IN_FEATURES: default feature dim for model init (will auto-rebuild if input F differs)
• MODEL_TYPE: mlp | graphsage
• CKPT_PATH: optional path inside repo, e.g. /repository/weights.pt

Local test

pip install -r requirements.txt
python inference.py --input-json example_node.json
python inference.py --input-json example_graph.json --model-type mlp

Upload to Hugging Face

See top-level instructions in your project README or follow the CLI snippet:

pip install -U huggingface_hub
huggingface-cli login
python - << 'PY'
from huggingface_hub import HfApi, upload_folder
api = HfApi()
REPO_ID = "YOUR_NS/elliptic-fraud"
api.create_repo(REPO_ID, repo_type="model", exist_ok=True)
upload_folder(repo_id=REPO_ID, folder_path=".", path_in_repo=".", commit_message="init graph-ready scaffold")
PY