baseline.py

"""Scalar baselines on the released splits. Run inside the downloaded repo."""
import json
from pathlib import Path
import numpy as np

if __name__=="__main__":
    root=Path(__file__).resolve().parent
    info=json.loads((root/"baselines.json").read_text()); target=info["target"]
    train=[json.loads(s) for s in (root/"train.jsonl").read_text().splitlines()]
    test=[json.loads(s) for s in (root/"test.jsonl").read_text().splitlines()]
    exclude={"id","split","hdf5","group","nodes","cells","parameter_sha256","relative_nodal_l2_error",target}
    names=[k for k in train[0] if k not in exclude]
    X=np.array([[r[k] for k in names] for r in train]); Xt=np.array([[r[k] for k in names] for r in test])
    y=np.array([r[target] for r in train]); yt=np.array([r[target] for r in test])
    mu=X.mean(0); s=X.std(0); X=(X-mu)/s; Xt=(Xt-mu)/s
    idx=np.argmin(((Xt[:,None,:]-X[None,:,:])**2).sum(2),axis=1)
    for name,pred in [("train_mean",np.full_like(yt,y.mean())),("nearest_neighbor",y[idx])]:
        print(name,"relative L2",np.linalg.norm(pred-yt)/np.linalg.norm(yt),"MAE",np.mean(abs(pred-yt)))