r"""diag.py — why co R² is -3.5 and pm2_5 flat. Checks prediction vs target scale per pollutant."""
import os,sys,json,pathlib,argparse
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"]
ap=argparse.ArgumentParser(); ap.add_argument("--tensors",default="data/tensors"); ap.add_argument("--ckpt",default="data/atmosiq_best.pt")
a=ap.parse_args(); dev="cuda" if torch.cuda.is_available() else "cpu"
meta=json.load(open(pathlib.Path(a.tensors)/"meta.json"))
data=AtmosData(a.tensors,window=7,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()
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=(data.split==2).float().unsqueeze(-1).to(dev)
# raw target ranges (test fold, labeled)
print("=== target ranges (test fold, raw concentration) ===")
for j,p in enumerate(POLL):
    m=(data.M[:,:,j]>0)&(data.split==2).unsqueeze(0)
    vals=data.Y[:,:,j][m].float()
    if len(vals): print(f"  {p:6}: min={vals.min():.2f} median={vals.median():.2f} max={vals.max():.2f} mean={vals.mean():.2f}")
# sample predictions on a few test days
print("\n=== predicted (expm1 mu) vs truth, horizon0, sample ===")
with torch.no_grad():
    for t in data.target_days(2)[:3]:
        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,:]; pred=(torch.exp(mu.clamp(max=12))-1e-3).clamp_min(0); y=b["y"][:,0,:]; m=(b["m"][:,0,:]*te)
        for j,p in enumerate(POLL):
            mm=m[:,j]>0
            if mm.sum()>2:
                print(f"  day{t} {p:6}: pred[med={pred[mm,j].median():.2f} max={pred[mm,j].max():.1f}] truth[med={y[mm,j].median():.2f}] mu[med={mu[mm,j].median():.2f}]")
        break