File size: 4,209 Bytes
5d0862f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | r"""
evaluate.py — M2 Go/No-Go. Loads atmosiq_best.pt, predicts on the TEST fold (spatial
held-out nodes, full-graph), reports per-pollutant spatial-held-out R²/RMSE/MAE at
horizon 0, and compares to the M1 XGBoost floor (refs.json). Blueprint §10 primary metric.
Run: python data\models\evaluate.py --tensors data\tensors --ckpt data\atmosiq_best.pt
"""
from __future__ import annotations
import os, sys, json, argparse, pathlib
import numpy as np, torch
sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent))
from model import AtmosIQNet
from dataset_v2 import AtmosData
POLL=["pm2_5","pm10","no2","so2","co","o3","nh3"]
def main():
ap=argparse.ArgumentParser()
ap.add_argument("--tensors",default="data/tensors"); ap.add_argument("--ckpt",default="data/atmosiq_best.pt")
ap.add_argument("--window",type=int,default=7); ap.add_argument("--refs",default="data/models/refs.json")
ap.add_argument("--test_days",type=int,default=0,help="0=all test days")
a=ap.parse_args()
dev="cuda" if torch.cuda.is_available() else "cpu"; print("device:",dev)
meta=json.load(open(pathlib.Path(a.tensors)/"meta.json"))
data=AtmosData(a.tensors, window=a.window, device=dev)
cfg={"calendar_num_dim":meta["dimsD_num"],"n_recon_channels":9+meta["dimsB"]+meta["dimsC_static"]+meta["dimsC_yearly"],
"categorical":{"land_use_class":4,"grap_stage":3,"n_landuse":meta["n_landuse"],"n_grap":meta["n_grap"]}}
net=AtmosIQNet(cfg).to(dev); net.load_state_dict(torch.load(a.ckpt,map_location=dev)); net.eval()
nz=np.load(pathlib.Path(a.tensors)/"norm.npz")
ylm=torch.tensor(nz["y_logmean"]).float().to(dev) if "y_logmean" in nz else None
yls=torch.tensor(nz["y_logstd"]).float().to(dev) if "y_logstd" in nz else None
print(f"loaded {a.ckpt} ({sum(p.numel() for p in net.parameters()):,} params)")
alln=torch.arange(data.N,device=data.store)
ei,eh,nh=data.ei.to(dev),data.e_hat.to(dev),data.n_hat.to(dev)
te_days=data.target_days(2)
if a.test_days: te_days=te_days[:a.test_days]
te_mask=(data.split==2).float().unsqueeze(-1).to(dev)
print(f"test days: {len(te_days)} | held-out test nodes: {int((data.split==2).sum())}")
# collect horizon-0 predictions (median = mu in log space -> expm1) vs truth on test nodes
Ys={p:[] for p in POLL}; Ps={p:[] for p in POLL}
with torch.no_grad():
for t in te_days:
b=data.window(t,alln,ei,eh,nh)
out=net(b["groups_t"],b["ei"],b["e_hat"],b["n_hat"],b["wu"],b["wv"],b["blh"],b["comp"])
mu=out["mu"][:,0,:] # horizon 0 [N,7] (standardized log space)
if ylm is not None: mu=mu*yls.view(1,-1)+ylm.view(1,-1) # de-standardize
pred=(torch.exp(mu.clamp(max=12))-1e-3).clamp_min(0) # log(y+eps) inverse; clamp>=0 (fixes co negatives)
y=b["y"][:,0,:]; m=(b["m"][:,0,:]*te_mask)
for j,p in enumerate(POLL):
mm=m[:,j]>0
if mm.any():
Ys[p].append(y[mm,j].cpu()); Ps[p].append(pred[mm,j].cpu())
refs=json.load(open(a.refs)) if os.path.exists(a.refs) else {}
print("\n===== M2 — SPATIAL-HELD-OUT R² (horizon 0, blueprint §10) =====")
print(f"{'pollutant':8} {'n':>6} {'R²':>8} {'RMSE':>8} {'MAE':>8} vs XGBoost")
r2s=[]
for p in POLL:
if not Ys[p]: print(f"{p:8} {'--':>6} no test labels"); continue
y=torch.cat(Ys[p]).numpy(); pr=torch.cat(Ps[p]).numpy()
pr=np.nan_to_num(pr,nan=np.nanmean(y))
ss_res=((y-pr)**2).sum(); ss_tot=((y-y.mean())**2).sum()+1e-9
r2=1-ss_res/ss_tot; rmse=np.sqrt(((y-pr)**2).mean()); mae=np.abs(y-pr).mean()
xg=refs.get(p,{}).get("xgboost")
cmp=f"xgb {xg:+.2f} {'BEAT' if (xg is not None and r2>xg) else 'below' if xg is not None else ''}"
print(f"{p:8} {len(y):>6} {r2:>+8.3f} {rmse:>8.2f} {mae:>8.2f} {cmp}")
if p not in ("nh3",): r2s.append(r2) # nh3 excluded from Go/No-Go (sparse-aux)
print(f"\nmean primary R² (excl nh3): {np.mean(r2s):+.3f}")
print("Go/No-Go: GNN should beat XGBoost floor on primary pollutants (esp pm2_5) on SPATIAL held-out.")
if __name__=="__main__": main() |