"""ARPACK dip-screening of a dip-farm trajectory (the governor's dataset): for every frozen snapshot >= --from_step, top-3 leading eigenvalues of the forward map at the 400-deep state. Output: one line per snapshot -> dips (all |lam|<1) marked DIP. Usage: dip_screen.py --seed N [--from_step 800]""" import argparse, numpy as np, torch, scipy.sparse.linalg as sla from pathlib import Path from torch.autograd.functional import jvp import lt_ep_train as L ap = argparse.ArgumentParser() ap.add_argument('--seed', type=int, required=True) ap.add_argument('--from_step', type=int, default=800) cfg = ap.parse_args() EPS, B, C = 0.1, 6, 1.0 traj = Path(f'runs/dipfarm_s{cfg.seed}_traj') steps = sorted(int(p.stem[1:]) for p in traj.glob('s*.pt') if int(p.stem[1:]) >= cfg.from_step) print(f"[seed {cfg.seed}] screening {len(steps)} snapshots: {steps}", flush=True) for s in steps: torch.manual_seed(0) blk = L.EQBlock(512, 16, 256, 256, c=C, attn_mode='thick'); blk.qknorm = True ck = torch.load(traj / f's{s}.pt', map_location=L.dev) with torch.no_grad(): for p, w in zip(blk.allp, ck['allp']): p.copy_(w.to(L.dev)) torch.manual_seed(42) idx, _ = L.get_batch('train', B, 256) xin = blk.embed(idx).detach() z = L.relax(blk, xin.clone(), xin, 400, EPS) sh, n = z.shape, z.numel() kk = 1.0 - EPS * (1.0 + C) def mv(x, z=z, sh=sh): v = torch.from_numpy(np.asarray(x, dtype=np.float32)).to(L.dev).view(sh) with torch.no_grad(): Mv = kk * v + EPS * jvp(blk.nc_force, z, v)[1] return Mv.reshape(-1).double().cpu().numpy() A = sla.LinearOperator((n, n), matvec=mv, dtype=np.float64) try: vals = sorted(sla.eigs(A, k=3, which='LM', return_eigenvectors=False, maxiter=1500, tol=2e-4), key=lambda x: -abs(x)) stable = all(abs(l) < 1.0 for l in vals) out = " ".join(f"|l|={abs(l):.5f}" for l in vals) print(f"s{cfg.seed}/{s:<5} {out} {'DIP' if stable else ''}", flush=True) except Exception as e: print(f"s{cfg.seed}/{s:<5} ARPACK-fail {type(e).__name__}", flush=True) print(f"DIP_SCREEN_S{cfg.seed}_DONE", flush=True)