# scripts/eval_multistep.py import os, sys, argparse, json from pathlib import Path from datetime import timedelta import torch import numpy as np import pandas as pd import xarray as xr from tqdm import tqdm from omegaconf import OmegaConf from hydra.utils import instantiate sys.path.append(str(Path(__file__).parent.parent)) from evaluation.deterministic_metrics import headline_wrmse from dataloaders.preprocessors import reverse_default_preprocessor # -------------------- helpers -------------------- def _time_from_sample(sample): t = sample.get("time") if isinstance(t, (list, tuple)): t = t[0] return t def _prev_offset(ds): # Era5Forecast.__len__ 보정과 동일 규칙 return (ds.lead_time_hours // ds.timedelta) if getattr(ds, "load_prev", False) else 0 def _time_map_fast(ds): """__getitem__ 호출 없이 id2pt에서 바로 map 구성 (매우 빠름).""" off = _prev_offset(ds) m = {} for i in range(len(ds)): _, _, ts = ds.id2pt[i + off] # (file_id, line_id, np.datetime64) t = pd.to_datetime(ts).strftime('%Y-%m-%d-%H-%M') m[t] = i return m def _latlon_from_hw(H, W): lat_step = 180.0 / H lon_step = 360.0 / W lats = np.arange(-90.0, 90.0, lat_step) # H개 lons = np.arange(0.0, 360.0, lon_step) # W개 return lats, lons def _area_weights(H, W, device, dtype): lats, _ = _latlon_from_hw(H, W) w = np.cos(np.deg2rad(lats)) w = w / w.mean() w = torch.as_tensor(w, device=device, dtype=dtype)[:, None].repeat(1, W) # [H,W] return w def _to_dt(tstr: str): return pd.to_datetime(tstr, utc=True, format="%Y-%m-%d-%H-%M").tz_localize(None) def _fmt_dt(dt) -> str: return dt.strftime("%Y-%m-%d-%H-%M") def _pair_hr_lr_at_time(t_str, hr_norm_ds, lr_norm_ds, t2i_hr, t2i_lr): """현재 t의 HR 정규배치 + LR(now, prev)의 surface 4ch를 붙여서 반환. prev LR가 없으면 None (스킵).""" if t_str not in t2i_hr or t_str not in t2i_lr: return None ih, il = t2i_hr[t_str], t2i_lr[t_str] a = hr_norm_ds[ih] # HR normalized b = lr_norm_ds[il] # LR normalized # prev LR 시각 lead = int(getattr(hr_norm_ds, "lead_time_hours", 24)) prev_t = _fmt_dt(_to_dt(t_str) - pd.Timedelta(hours=lead)) if prev_t not in t2i_lr: return None b_prev = lr_norm_ds[t2i_lr[prev_t]] if "state_surface" not in b or "state_surface" not in b_prev: return None out = dict(a) out["state_surface_lr"] = b["state_surface"] out["prev_state_surface_lr"] = b_prev["state_surface"] # time 표준화 t = _time_from_sample(a) out["time"] = [t] if isinstance(t, str) else a["time"] return out def _merge_next_pair_from_pred(hr_norm_ds, lr_norm_ds, t2i_lr, pred, hr_norm_batch): """HR는 pred 기반으로 unroll(normalize_next_batch), LR는 실측 now/prev로 갱신. 두 LR 타임이 없으면 None.""" hr_next = hr_norm_ds.normalize_next_batch(pred, hr_norm_batch) next_t = hr_next["time"][0] if isinstance(hr_next["time"], (list, tuple)) else hr_next["time"] # next 시점의 now/prev LR 확보 lead = int(getattr(hr_norm_ds, "lead_time_hours", 24)) prev_t = _fmt_dt(_to_dt(next_t) - pd.Timedelta(hours=lead)) if next_t not in t2i_lr or prev_t not in t2i_lr: return None b_now = lr_norm_ds[t2i_lr[next_t]] b_prev = lr_norm_ds[t2i_lr[prev_t]] if "state_surface" not in b_now or "state_surface" not in b_prev: return None out = dict(hr_next) out["state_surface_lr"] = b_now["state_surface"] out["prev_state_surface_lr"] = b_prev["state_surface"] return out def _build_model(cfg): """ 다양한 cfg 형태를 지원해서 backbone/module을 생성. - 형태1: cfg.module.module + cfg.module.backbone - 형태2: cfg.module(_target_ 포함) + cfg.backbone - 형태3: cfg.module(_target_) 내부에서 backbone까지 생성 """ if "module" not in cfg: raise RuntimeError("cfg.module 노드가 없습니다. --cfg-path를 확인하세요.") node = cfg.module # 형태1 if (isinstance(node, dict) or ("_target_" not in node)) and ("module" in node and "backbone" in node): backbone = instantiate(node.backbone) pl_module = instantiate(node.module, backbone=backbone) return pl_module # 형태2 if "_target_" in node and "backbone" in cfg: backbone = instantiate(cfg.backbone) pl_module = instantiate(node, backbone=backbone) return pl_module # 형태3 if "_target_" in node: pl_module = instantiate(node) return pl_module raise RuntimeError("알 수 없는 cfg 구조입니다. cfg.module을 확인하세요.") # -------------------- main -------------------- def main(): ap = argparse.ArgumentParser() ap.add_argument("--ckpt-path", type=str, required=True) ap.add_argument("--cfg-path", type=str, required=True) ap.add_argument("--dataset-cfg-path", type=str, required=True) ap.add_argument("--multistep", type=int, default=10) ap.add_argument("--force", action="store_true") ap.add_argument("--debug", action="store_true") args = ap.parse_args() torch.set_grad_enabled(False) device = "cuda:0" if torch.cuda.is_available() else "cpu" write_frequency = 10 # ---------- load cfgs ---------- cfg = OmegaConf.load(args.cfg_path) ds_cfg = OmegaConf.load(args.dataset_cfg_path) # ---------- build model & load ckpt ---------- pl_module = _build_model(cfg) state = torch.load(args.ckpt_path, map_location="cpu", weights_only=False) state_dict = state.get("state_dict", state) pl_module.load_state_dict(state_dict, strict=False) pl_module = pl_module.to(device).eval() if hasattr(pl_module, "use_prev"): pl_module.use_prev = True # prev LR/HR 경로 활성화 # ---------- resolve dataset config ---------- # 허용: # A) {dataloader: {dataset_hr: ..., dataset_lr: ...}} # B) {dataset_hr: ..., dataset_lr: ...} # C) {dataset: ...} (HR/LR 동일 사용) if "dataloader" in ds_cfg: dl_cfg = ds_cfg.dataloader else: dl_cfg = ds_cfg if "dataset_hr" in dl_cfg and "dataset_lr" in dl_cfg: ds_hr_node = dl_cfg.dataset_hr ds_lr_node = dl_cfg.dataset_lr elif "dataset" in dl_cfg: ds_hr_node = dl_cfg.dataset ds_lr_node = dl_cfg.dataset else: raise RuntimeError( "dataset config 형태가 다릅니다. 다음 중 하나여야 합니다:\n" " - dataloader.dataset_hr & dataloader.dataset_lr\n" " - dataset_hr & dataset_lr\n" " - dataset" ) # ---------- datasets (정규화/원시) ---------- # 정규화된 HR/LR (모델 입력용) hr_norm_ds = instantiate(ds_hr_node, domain="test", z0012=True) lr_norm_ds = instantiate(ds_lr_node, domain="test", z0012=True) # 원시 HR (denorm GT 저장/메트릭 계산용) hr_raw_ds = instantiate( ds_hr_node, domain="test", z0012=True, multistep=args.multistep, input_norm_scheme=None, output_norm_scheme=None, ) # 시간 매핑(빠른 버전) t2i_hr = _time_map_fast(hr_norm_ds) t2i_lr = _time_map_fast(lr_norm_ds) raw_loader = torch.utils.data.DataLoader(hr_raw_ds, batch_size=1, num_workers=4, shuffle=False) outs = dict(traj_surface=[], traj_level=[], timestamps=[]) err_log = [] lvl_names = hr_norm_ds.variables["state_level"] # ex) ['Z50',...,'Q1000'] srf_names = hr_norm_ds.variables["state_surface"] # ['U10','V10','T2m','SP'] # ---- per-var/time accumulators for WRMSE(JSON) ---- want_surf = ["U10", "V10", "T2m", "SP"] want_lvl = {"Z500":("Z",500), "U850":("U",850), "V850":("V",850), "T850":("T",850), "Q700":("Q",700)} pres_list = hr_norm_ds.pressure_levels # [50,...,1000] pres2idx = {int(p): i for i, p in enumerate(pres_list)} var2idx = {"Z":0, "U":1, "V":2, "T":3, "Q":4} json_sums, json_den = {}, {} def accumulate_json(key, err2_map, wmap): json_sums[key] = json_sums.get(key, 0.0) + float((err2_map * wmap).sum().item()) json_den[key] = json_den.get(key, 0.0) + float(wmap.sum().item()) first_debug_printed = False ckpt_stem = Path(args.ckpt_path).stem save_dir = Path("evalstore") / ckpt_stem save_dir.mkdir(parents=True, exist_ok=True) zarr_path = save_dir / f"test2020-multistep{args.multistep}.zarr" if zarr_path.exists() and args.force: import shutil; shutil.rmtree(zarr_path) # ---------- per-time rollout ---------- debug_steps = {1, 2, 10} # 1-based printed_steps = set() first_valid_seen = False for i, raw in tqdm(enumerate(raw_loader), total=len(raw_loader)): t0 = _time_from_sample(raw) pair = _pair_hr_lr_at_time(t0, hr_norm_ds, lr_norm_ds, t2i_hr, t2i_lr) if pair is None: continue # prev LR가 없는 시작 시점은 스킵 is_first_valid = False if not first_valid_seen: is_first_valid = True first_valid_seen = True hr_batch = {k: v for k, v in pair.items() if not k.endswith("_lr")} traj_s, traj_l = [], [] for k in range(args.multistep): pair_k = {kk: (vv.to(device) if hasattr(vv, "to") else vv) for kk, vv in pair.items()} pred = pl_module.forward(pair_k) # 디바이스 혼합 방지: denorm 계산은 CPU에서 denorm_pred = hr_norm_ds.denormalize({k2: (v2.detach().cpu() if torch.is_tensor(v2) else v2) for k2, v2 in pred.items()}, {k2: (v2.cpu() if torch.is_tensor(v2) else v2) for k2, v2 in hr_batch.items()}) traj_s.append(denorm_pred["next_state_surface"].cpu()) traj_l.append(denorm_pred["next_state_level"].cpu()) # ---- per-var/time JSON 누적 ---- s_pred = denorm_pred["next_state_surface"].squeeze(-3) # [B,4,H,W] l_pred = denorm_pred["next_state_level"] # [B,5,L,H,W] if k == 0: s_gt = raw["next_state_surface"].squeeze(-3) # [B,4,H,W] l_gt = raw["next_state_level"] # [B,5,L,H,W] else: s_gt = raw["future_state_surface"][:, k-1].squeeze(-3) l_gt = raw["future_state_level"][:, k-1] B, _, H, W = s_pred.shape wmap = _area_weights(H, W, device=s_pred.device, dtype=s_pred.dtype) # [H,W] # surface vars for vi, name in enumerate(want_surf): err2 = (s_pred[:, vi] - s_gt[:, vi]).pow(2).mean(0) # [H,W] key = f"lr2hr_H{H}xW{W}/t{k+1}/{name}" accumulate_json(key, err2, wmap) # level vars at target pressures for jname, (vtag, plevel) in want_lvl.items(): vc = var2idx[vtag] li = pres2idx[int(plevel)] err2 = (l_pred[:, vc, li] - l_gt[:, vc, li]).pow(2).mean(0) key = f"lr2hr_H{H}xW{W}/t{k+1}/{jname}" accumulate_json(key, err2, wmap) # ---- 첫 유효 스텝에서 디버그 프린트 (항상 1회) ---- if is_first_valid and ((k+1) in debug_steps) and ((k+1) not in printed_steps): print(f"\n[DEBUG — first valid rollout @ t={k+1}]") for name in ["state_surface_lr","prev_state_surface_lr","state_constant","state_level"]: if name in pair_k and hasattr(pair_k[name], "shape"): print(f" batch['{name}']:", tuple(pair_k[name].shape)) print(" pred.next_state_surface:", tuple(denorm_pred["next_state_surface"].shape)) print(" pred.next_state_level: ", tuple(denorm_pred["next_state_level"].shape)) # 간단 WRMSE 샘플 def wrmse_scalar(a, b, w): e2 = (a - b).pow(2).mean(0) # [H,W] return float(torch.sqrt((e2 * w).sum() / w.sum()).item()) u10_rmse = wrmse_scalar(s_pred[:,0], s_gt[:,0], wmap) z500_rmse= wrmse_scalar(l_pred[:,0,pres2idx[500]], l_gt[:,0,pres2idx[500]], wmap) print(f" U10@t{k+1} WRMSE: {u10_rmse:.4f}") print(f" Z500@t{k+1} WRMSE: {z500_rmse:.4f}\n") printed_steps.add(k+1) # 다음 스텝용 페어 갱신 (HR unroll + 실측 LR now/prev) merged = _merge_next_pair_from_pred(hr_norm_ds, lr_norm_ds, t2i_lr, pred, hr_batch) if merged is None: break pair = merged hr_batch = {k2: v2 for k2, v2 in pair.items() if not k2.endswith("_lr")} if not traj_s: continue # 안전장치 traj_s = torch.stack(traj_s, dim=1) # [1,T,4,1,H,W] 또는 [1,T,4,H,W] traj_l = torch.stack(traj_l, dim=1) # [1,T,5,L,H,W] outs["timestamps"].append(t0) outs["traj_surface"].append(traj_s) outs["traj_level"].append(traj_l) # headline metrics gt_traj = dict( traj_surface=torch.cat([raw["next_state_surface"][:, None], raw["future_state_surface"]], dim=1), traj_level =torch.cat([raw["next_state_level"][:, None], raw["future_state_level"]], dim=1), ) err_log.append(headline_wrmse(dict(traj_surface=traj_s, traj_level=traj_l), gt_traj, prefix="traj")) # 주기적 저장 if (i % write_frequency == 0) or (i == len(raw_loader) - 1): level_outs = torch.cat(outs["traj_level"]).cpu() level_outs = reverse_default_preprocessor(level_outs).numpy() surface_outs = torch.cat(outs["traj_surface"]).cpu() surface_outs = reverse_default_preprocessor(surface_outs).numpy() Hs, Ws = surface_outs.shape[-3], surface_outs.shape[-2] lats, lons = _latlon_from_hw(Hs, Ws) T = surface_outs.shape[1] xds = xr.Dataset( data_vars={ **{lvl_names[ii]: (["time", "prediction_timedelta", "level", "latitude", "longitude"], level_outs[:, :, ii]) for ii in range(len(lvl_names))}, **{srf_names[ii]: (["time", "prediction_timedelta", "latitude", "longitude"], surface_outs[:, :, ii, 0]) for ii in range(len(srf_names))}, }, coords=dict( time=pd.to_datetime(outs["timestamps"], utc=True, format="%Y-%m-%d-%H-%M").tz_localize(None), prediction_timedelta=[timedelta(k+1) for k in range(T)], latitude=lats, longitude=lons, level=[50,100,150,200,250,300,400,500,600,700,850,925,1000], ), ) xds.to_zarr( zarr_path, append_dim="time" if i else None, encoding=dict(time=dict(units="hours since 2000-01-01")) if not i else None, ) # running errors if err_log: avg_err = {k: torch.cat([e[k] for e in err_log], dim=0).mean(0) for k in err_log[0].keys()} print("[running headline_wrmse]", {k: (v.item() if torch.numel(v)==1 else v.mean().item()) for k, v in avg_err.items()}) outs = dict(traj_surface=[], traj_level=[], timestamps=[]) if args.debug: break # 최종 headline 저장 if err_log: avg_err = {k: torch.cat([e[k] for e in err_log], dim=0).mean(0) for k in err_log[0].keys()} torch.save(avg_err, save_dir / f"allmetrics-multistep{args.multistep}.pt") print("[final headline_wrmse]", {k: (v.item() if torch.numel(v)==1 else v.mean().item()) for k, v in avg_err.items()}) # ---- JSON 출력 (per-var/time WRMSE) ---- json_out = {} for key in sorted(json_sums.keys()): rmse = np.sqrt(json_sums[key] / max(json_den[key], 1e-12)) json_out[key] = float(rmse) json_path = save_dir / f"pervar_rmse_t{args.multistep}.json" with open(json_path, "w") as f: json.dump(json_out, f, indent=2) print(f"[saved JSON] {json_path}") if __name__ == "__main__": main()