code/ezflow_v3/gnn/plot_ood.py

"""CPU-only actual-vs-predicted field plots for OOD cases (uses best.pt).

Renders, for a few held-out (OOD) cases, a z~0 mid-plane slice colored by the
true CFD field, the model prediction, and the absolute error -- for both velocity
magnitude |U| and pressure p. Pure CPU; safe to run alongside GPU training.

Run:  python -m ezflow_v3.gnn.plot_ood --n 2
"""
from __future__ import annotations
import argparse, os, sys
import numpy as np
import torch
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import matplotlib.tri as mtri

os.environ.setdefault("KMP_DUPLICATE_LIB_OK", "TRUE")
os.environ.setdefault("OMP_NUM_THREADS", "2")
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
from ezflow_v3.gnn.etl import CaseDatasetV2
from ezflow_v3.gnn.model_v5 import MeshGraphNetV5
from ezflow_v3.gnn import features as F
from ezflow_v3.gnn.train_v5 import split


def r2(t, p):
    ss = ((t - p) ** 2).sum(); tot = ((t - t.mean()) ** 2).sum() + 1e-30
    return float(1 - ss / tot)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--run", default=os.path.join(os.path.dirname(__file__), "_runs", "rans_v5"))
    ap.add_argument("--cache", default=r"C:\dev\ezflow_eval\cache_v3")
    ap.add_argument("--out", default=r"C:\dev\ezflow_eval\ood_plots")
    ap.add_argument("--ckpt", default="best.pt")
    ap.add_argument("--n", type=int, default=2)
    a = ap.parse_args()
    os.makedirs(a.out, exist_ok=True)

    ck = torch.load(os.path.join(a.run, a.ckpt), map_location="cpu", weights_only=False)
    model = MeshGraphNetV5(F.NODE_FEATURE_DIM, 4, F.GLOBAL_DIM,
                           hidden=int(ck["args"]["hidden"]), K=int(ck["args"]["K"]),
                           out_dim=F.TARGET_DIM)
    model.load_state_dict(ck["model"]); model.eval()
    ep = ck.get("epoch", "?")
    print(f"loaded {a.ckpt} (epoch {ep}, val_loss {ck.get('val_loss','?')})", flush=True)

    nz = np.load(os.path.join(a.run, "norms.npz"))
    ym = torch.tensor(nz["y_mean"]); ys = torch.tensor(nz["y_std"])
    gm = torch.tensor(nz["g_mean"]); gs = torch.tensor(nz["g_std"])

    ds = CaseDatasetV2(a.cache)
    _, _, ood = split(ds)

    for setname, graphs in ood.items():
        sel = graphs[:a.n]
        if not sel:
            continue
        fig, axes = plt.subplots(len(sel), 6, figsize=(23, 3.6 * len(sel) + 0.8), squeeze=False)
        for r, g in enumerate(sel):
            gg = g.clone(); gg.global_feat = (gg.global_feat - gm) / gs
            with torch.no_grad():
                pred = model(gg).numpy() * ys.numpy() + ym.numpy()   # -> physical
            gt = g.y.numpy()
            pos = g.pos.numpy()
            uT = np.linalg.norm(gt[:, :3], axis=1); uP = np.linalg.norm(pred[:, :3], axis=1)
            pT, pP = gt[:, 3], pred[:, 3]
            r2u, r2p = r2(uT, uP), r2(pT, pP)          # R2 on the full graph
            Re = float(getattr(g, "Re", 0.0))

            # z~0 mid-plane slice (flow is in x-y), then crop to body + wake
            z = pos[:, 2]; tol = 0.15
            while (np.abs(z) < tol).sum() < 800 and tol < 1.0:
                tol += 0.08
            body = uT < 0.2
            cx = np.median(pos[body, 0]) if body.any() else 0.0
            cy = np.median(pos[body, 1]) if body.any() else 0.0
            m = (np.abs(z) < tol) & (pos[:, 0] > cx - 2.5) & (pos[:, 0] < cx + 8.0) \
                & (np.abs(pos[:, 1] - cy) < 3.0)
            X, Y = pos[m, 0], pos[m, 1]

            # triangulate the slice, mask triangles bridging the body/far gaps
            tri = mtri.Triangulation(X, Y)
            t = tri.triangles
            e = np.concatenate([np.hypot(X[t[:, i]] - X[t[:, j]], Y[t[:, i]] - Y[t[:, j]])
                                for i, j in ((0, 1), (1, 2), (2, 0))])
            maxe = np.maximum.reduce([np.hypot(X[t[:, 0]] - X[t[:, 1]], Y[t[:, 0]] - Y[t[:, 1]]),
                                      np.hypot(X[t[:, 1]] - X[t[:, 2]], Y[t[:, 1]] - Y[t[:, 2]]),
                                      np.hypot(X[t[:, 2]] - X[t[:, 0]], Y[t[:, 2]] - Y[t[:, 0]])])
            tri.set_mask(maxe > 4.5 * np.median(e))
            xlo, xhi = np.percentile(X, 1) - 0.3, np.percentile(X, 99) + 0.3
            ylo, yhi = np.percentile(Y, 1) - 0.3, np.percentile(Y, 99) + 0.3

            def panel(c, val, title, vmin, vmax, cmap):
                lv = np.linspace(vmin, vmax, 25)
                cs = axes[r, c].tricontourf(tri, val[m], levels=lv, cmap=cmap, extend="both")
                axes[r, c].set_title(title, fontsize=9)
                axes[r, c].set_aspect("equal"); axes[r, c].set_xticks([]); axes[r, c].set_yticks([])
                axes[r, c].set_xlim(xlo, xhi); axes[r, c].set_ylim(ylo, yhi)
                plt.colorbar(cs, ax=axes[r, c], fraction=0.046, pad=0.02)

            vU = (float(min(uT.min(), uP.min())), float(max(uT.max(), uP.max())))
            vP = (float(min(pT.min(), pP.min())), float(max(pT.max(), pP.max())))
            eU = np.abs(uP - uT); eP = np.abs(pP - pT)
            panel(0, uT, f"{g.cid}\nRe={Re:.0f}  |U| TRUE", *vU, "viridis")
            panel(1, uP, f"|U| PRED  (R2={r2u:.3f})", *vU, "viridis")
            panel(2, eU, "|U| |error|", 0.0, float(eU[m].max() + 1e-9), "magma")
            panel(3, pT, "p TRUE", *vP, "coolwarm")
            panel(4, pP, f"p PRED  (R2={r2p:.3f})", *vP, "coolwarm")
            panel(5, eP, "p |error|", 0.0, float(eP[m].max() + 1e-9), "magma")
            print(f"  {setname:12s} {g.cid:24s} Re={Re:7.0f}  |U|R2={r2u:.3f}  pR2={r2p:.3f}  (slice n={int(m.sum())}, tol={tol:.2f})", flush=True)

        fig.suptitle(f"OOD: {setname} - actual vs predicted ({a.ckpt} ep{ep}, z~0 slice)", fontsize=13)
        fig.tight_layout(rect=[0, 0, 1, 0.96])
        outp = os.path.join(a.out, f"ood_{setname}.png")
        fig.savefig(outp, dpi=110); plt.close(fig)
        print(f"saved {outp}", flush=True)
    print("PLOT_OOD_DONE", flush=True)


if __name__ == "__main__":
    main()