Blame for experiments/47_final_ens.py

"""Final four-modality evaluation with multi-encoder late fusion (frozen UNI2 anchor; real pairs only; virtual /
synthetic data only at test time). Combines the winners of 44_rnafm_eval.py (RNA) and 46_text_enc_eval.py (text).

Every head maps one frozen encoder into the frozen UNI2 space (S4 recipe of 42_: MLP for mIF, linear for RNA / text), so
the heads of several encoders of one modality can be averaged at test time:  z_mod = normalise(sum_e normalise(head_e(x_e))).
Configurations (mIF = VirTues 16-ch + MLP head in all of them):
  S4        RNA hvg                                   text BiomedCLIP                  (current method)
  single    RNA AIDO.Cell                             text KEEP                        (best single encoders)
  ens-small RNA hvg + Tahoe-x1 + SE-600M + AIDO.Cell  text KEEP + BiomedCLIP + Qwen3-Emb-0.6B
  ens-large RNA all 10 encoders                       text KEEP + BiomedCLIP + Qwen3-0.6B + PubMedBERT-emb + MedCPT
Pairs per fold k (same splits / test items as 42_ / 44_ / 46_): trained H&E<->mIF (ORION), H&E<->RNA (HEST),
H&E<->text (TCGA); emergent mIF<->RNA(v), mIF<->text(s) (ORION), text<->RNA(v), text<->mIF(MIPHEI) (TCGA);
external DLPFC (RNA heads retrained without spatialLIBD, inputs z-scored per dataset): H&E<->RNA, RNA<->text(s),
RNA<->mIF(MIPHEI).
Output results/final_ens_long.csv, results/final_ens_summary.csv, results/final_ens_tests.csv (paired vs S4)
"""
import os

import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
from scipy import stats
from scipy.optimize import linear_sum_assignment

from common import CACHE, ROOT, SPLITS

dev = "cuda"
H, RF = os.path.join(CACHE, "hest"), os.path.join(CACHE, "rnafm")
G100, DRAWS, BS, LR, EVAL_EVERY = 100, 50, 256, 3e-4, 50
f32 = lambda a: torch.from_numpy(np.asarray(a, dtype=np.float32))
nz = lambda x: F.normalize(x, dim=-1)
RNA_ALL = ["hvg", "tahoex1", "se600m", "scgpt", "nicheformer", "aidocell", "geneformer", "c2s", "genept-bge", "genept-bmc"]
CONF = {"S4": (["hvg"], ["biomedclip"]),
        "single": (["aidocell"], ["keep"]),
        "ens-small": (["hvg", "tahoex1", "se600m", "aidocell"], ["keep", "biomedclip", "qwen3-0.6b"]),
        "ens-large": (RNA_ALL, ["keep", "biomedclip", "qwen3-0.6b", "pubmedbert", "medcpt"])}

# ---------------------------------------------------------------- data
HE_O = f32(np.load(os.path.join(CACHE, "he_uni2.npy"))).to(dev)
IF_O = f32(np.load(os.path.join(CACHE, "if_virtues_m16.npy"))).to(dev)
IF_V_T = f32(np.load(os.path.join(CACHE, "if_virtues_m16_virtual_miphei_tcga_text.npy"))[:, 0]).to(dev)
IF_V_D = f32(np.load(os.path.join(CACHE, "if_virtues_m16_virtual_miphei_dlpfc.npy"))[:, 0]).to(dev)
HE_H = f32(np.load(os.path.join(H, "he_uni2.npy"))[:, 0]).to(dev)
HE_T = f32(np.load(os.path.join(CACHE, "allimg_he_uni2.npy"))).to(dev)
HE_D = f32(np.load(os.path.join(CACHE, "external", "dlpfc_he_uni2.npy"))).to(dev)
T = pd.read_csv(os.path.join(CACHE, "alltxt_meta.csv"))
T_IMG = torch.as_tensor(T.img_idx.values, device=dev)
T = T[~pd.read_csv(os.path.join(CACHE, "alltxt_junk.csv")).junk.values]
HM = pd.read_csv(os.path.join(H, "meta.csv"))
hvg = pd.read_csv(os.path.join(H, "hvg_genes.csv")).gene.values


def hvg_inputs():  # = 44_
    def virt(st):
        g = pd.read_csv(os.path.join(CACHE, "stpath", f"{st}_genes.csv")).gene.values
        c = np.expm1(np.load(os.path.join(CACHE, "stpath", f"{st}_pred.npy")).astype(np.float32)).clip(0)
        c = np.log1p(c / c.sum(1, keepdims=True) * 1e4)
        return c[:, pd.Series(range(len(g)), index=g).loc[hvg].values]
    eg = pd.read_csv(os.path.join(CACHE, "external", "dlpfc_genes_full.csv")).gene.values
    ec = np.load(os.path.join(CACHE, "external", "dlpfc_counts_full.npy")).astype(np.float32)
    ei = pd.Series(range(len(eg)), index=eg); ei = ei[~ei.index.duplicated()]
    d = np.log1p(ec / ec.sum(1, keepdims=True).clip(1) * 1e4)
    d = np.stack([d[:, ei[g]] if g in ei.index else np.zeros(len(d), np.float32) for g in hvg], 1)
    return {"hest": np.load(os.path.join(H, "hvg.npy")), "orion_v": virt("orion"), "tcga_v": virt("tcga_text"), "dlpfc": d}


RNA_IN = {e: ({s: f32(x).to(dev) for s, x in hvg_inputs().items()} if e == "hvg" else
              {s: f32(np.load(os.path.join(RF, f"{e}_{s}.npy"))).to(dev) for s in ["hest", "orion_v", "tcga_v", "dlpfc"]})
          for e in RNA_ALL}
TXT_ENCS = sorted({e for _, x in CONF.values() for e in x})
TXT_IN = {e: {s: f32(np.load(os.path.join(CACHE, f"{s}_text_{e}.npy"))[:, 0]).to(dev)
              for s in ["alltxt", "orion_gentext", "dlpfc_gentext"]} for e in TXT_ENCS}


def mlp(din, dout=1536, hid=512):
    return nn.Sequential(nn.Linear(din, hid), nn.GELU(), nn.Dropout(0.1), nn.Linear(hid, hid), nn.GELU(),
                         nn.LayerNorm(hid), nn.Linear(hid, dout))


def one_text_per_image(t_ids, rng):
    pick = T.loc[t_ids].groupby("img_idx").apply(lambda g: g.index[rng.integers(len(g))], include_groups=False)
    return pick.index.values, pick.values


def g100(a, b, seed=0):  # = 33_ / 42_
    a, b = a.cpu(), b.cpu()
    rng = np.random.default_rng(seed)
    out = {m: [] for m in ["R@1", "R@5", "R@10", "MRR", "pair_acc"]}
    for _ in range(DRAWS):
        s = torch.from_numpy(rng.choice(len(a), min(G100, len(a)), replace=False))
        sim = (a[s] @ b[s].T).numpy(); d = np.diag(sim)
        for S in (sim, sim.T):
            rank = (S > d[:, None]).sum(1)
            for k_ in (1, 5, 10):
                out[f"R@{k_}"].append((rank < k_).mean() / 2)
            out["MRR"].append((1 / (rank + 1)).mean() / 2)
        ri, ci = linear_sum_assignment(-sim)
        out["pair_acc"].append((ci[np.argsort(ri)] == np.arange(len(s))).mean())
    return {m: float(np.sum(v) / DRAWS) for m, v in out.items()}


def r1(a, b):
    s = a @ b.T
    return float((((s > s.diag()[:, None]).sum(1) == 0).float().mean() + ((s.T > s.diag()[:, None]).sum(1) == 0).float().mean()) / 2)


def train_head(X_tr, A_tr, X_va, A_va, kind, steps, seed, pos_tr=None):  # = 44_ / 46_
    torch.manual_seed(seed); rng = np.random.default_rng(seed)
    h = (mlp(X_tr.shape[1]) if kind == "mlp" else nn.Linear(X_tr.shape[1], 1536)).to(dev)
    logit = nn.Parameter(torch.tensor(np.log(1 / 0.07), dtype=torch.float32, device=dev))
    opt = torch.optim.AdamW(list(h.parameters()) + [logit], lr=LR, weight_decay=0.05)
    sch = torch.optim.lr_scheduler.OneCycleLR(opt, max_lr=LR, total_steps=steps, pct_start=0.05)
    za_va = nz(A_va[:, 0] if A_va.dim() == 3 else A_va)
    best, state = -1, None
    for step in range(1, steps + 1):
        h.train()
        b = torch.as_tensor(rng.choice(len(X_tr), BS, replace=False), device=dev)
        a = A_tr[b, torch.randint(0, A_tr.shape[1], (BS,), device=dev)] if A_tr.dim() == 3 else A_tr[b]
        lg = logit.exp().clamp(max=100) * nz(a) @ nz(h(X_tr[b])).T
        if pos_tr is None:
            t = torch.arange(BS, device=dev)
            loss = (F.cross_entropy(lg, t) + F.cross_entropy(lg.T, t)) / 2
        else:
            ii = pos_tr[b]; pos = (ii[:, None] == ii[None, :]).float(); tgt = pos / pos.sum(1, keepdim=True)
            loss = (torch.sum(-tgt * F.log_softmax(lg, 1), 1).mean() + torch.sum(-tgt * F.log_softmax(lg.T, 1), 1).mean()) / 2
        opt.zero_grad(set_to_none=True); loss.backward(); opt.step(); sch.step()
        if step % EVAL_EVERY == 0 or step == steps:
            h.eval()
            with torch.no_grad():
                v = r1(za_va, nz(h(X_va)))
            if v > best:
                best, state = v, {k_: x.detach().clone() for k_, x in h.state_dict().items()}
    h.load_state_dict(state); h.eval()
    return h


def zs(X, ref_set, idx, raw, ext=None):
    """z-score every set with the training rows of ref_set (raw = keep the baseline input unchanged, as in 42_);
    ext: sets z-scored on themselves (external data, = 38_ / 44_)"""
    if raw:
        out = dict(X)
    else:
        mu, sd = X[ref_set][idx].mean(0), X[ref_set][idx].std(0) + 1e-6
        out = {s: (v - mu) / sd for s, v in X.items()}
    for s in ext or []:
        out[s] = (X[s] - X[s].mean(0)) / (X[s].std(0) + 1e-6)
    return out


rows = []
for k in range(5):
    s = pd.read_csv(os.path.join(SPLITS, "patient_5fold", f"fold{k}.csv"))
    o = {x: s.tile_id[s.split == x].values for x in ["train", "val", "test"]}
    nif = lambda x: (x - IF_O[o["train"]].mean((0, 1))) / (IF_O[o["train"]].std((0, 1)) + 1e-3)
    h = pd.read_csv(os.path.join(SPLITS, "final", f"hest_fold{k}.csv"))
    hs = {x: h.spot_idx[h.split == x].values for x in ["train", "val", "test"]}
    sp = pd.read_csv(os.path.join(SPLITS, "final", f"tcga_fold{k}.csv")).set_index("img_idx").split
    tsp = sp.loc[T.img_idx].values
    ts = {x: T.index[tsp == x].values for x in ["train", "val", "test"]}
    vi, vt = (torch.as_tensor(x, device=dev) for x in one_text_per_image(ts["val"], np.random.default_rng(0)))
    ii, tt = (torch.as_tensor(x, device=dev) for x in one_text_per_image(ts["test"], np.random.default_rng(k)))
    trt = torch.as_tensor(ts["train"], device=dev)
    pool = np.flatnonzero(~HM.group.str.startswith("spatialLIBD").values)
    grp = np.random.default_rng(k).permutation(sorted(HM.group.iloc[pool].unique()))
    isva = HM.group.iloc[pool].isin(set(grp[: max(1, len(grp) // 10)])).values
    d_tr, d_va = pool[~isva], pool[isva]

    hm = mlp(512).to(dev)
    hm.load_state_dict(torch.load(os.path.join(ROOT, "runs", "rnafm", f"other_fold{k}.pt"), map_location=dev)["mif"])
    hm.eval()
    with torch.no_grad():
        Zm = {"he_o": nz(HE_O[o["test"], 0]), "orion": nz(hm(nif(IF_O[o["test"], 0]))),
              "tcga_v": nz(hm(nif(IF_V_T[ii]))), "dlpfc_v": nz(hm(nif(IF_V_D)))}
    Zr, Zt = {}, {}
    for e in RNA_ALL:
        X = zs(RNA_IN[e], "hest", hs["train"], e == "hvg")
        hr = train_head(X["hest"][hs["train"]], HE_H[hs["train"]], X["hest"][hs["val"]], HE_H[hs["val"]], "linear", 2700, k)
        Xd = zs(RNA_IN[e], "hest", d_tr, False, ext=["dlpfc"])
        hd = train_head(Xd["hest"][d_tr], HE_H[d_tr], Xd["hest"][d_va], HE_H[d_va], "linear", 2700, k)
        with torch.no_grad():
            Zr[e] = {"hest": nz(hr(X["hest"][hs["test"]])), "orion_v": nz(hr(X["orion_v"][o["test"]])),
                     "tcga_v": nz(hr(X["tcga_v"][ii])), "dlpfc": nz(hd(Xd["dlpfc"]))}
        print(f"fold {k} RNA head {e} done", flush=True)
    for e in TXT_ENCS:
        X = zs(TXT_IN[e], "alltxt", trt, e == "biomedclip")
        ht = train_head(X["alltxt"][trt], HE_T[T_IMG[trt]], X["alltxt"][vt], HE_T[vi, 0], "linear", 1000, k, pos_tr=T_IMG[trt])
        # DLPFC synthetic captions: z-scored with TCGA-train stats (text domain is the same caption style)
        with torch.no_grad():
            Zt[e] = {"tcga": nz(ht(X["alltxt"][tt])), "orion_s": nz(ht(X["orion_gentext"][o["test"]])),
                     "dlpfc_s": nz(ht(X["dlpfc_gentext"]))}
    he_h, he_t, he_d = nz(HE_H[hs["test"]]), nz(HE_T[ii, 0]), nz(HE_D)
    for cname, (re_, te_) in CONF.items():
        R = {s_: nz(sum(Zr[e][s_] for e in re_)) for s_ in Zr[re_[0]]}
        Tx = {s_: nz(sum(Zt[e][s_] for e in te_)) for s_ in Zt[te_[0]]}
        res = {("H&E<->mIF (ORION)", "trained"): g100(Zm["he_o"], Zm["orion"]),
               ("H&E<->RNA (HEST)", "trained"): g100(he_h, R["hest"]),
               ("H&E<->text (TCGA)", "trained"): g100(he_t, Tx["tcga"]),
               ("mIF<->RNA(virtual) (ORION)", "emergent"): g100(Zm["orion"], R["orion_v"]),
               ("mIF<->text(synthetic) (ORION)", "emergent"): g100(Zm["orion"], Tx["orion_s"]),
               ("text<->RNA(virtual) (TCGA)", "emergent"): g100(Tx["tcga"], R["tcga_v"]),
               ("text<->mIF(MIPHEI) (TCGA)", "emergent"): g100(Tx["tcga"], Zm["tcga_v"]),
               ("H&E<->RNA (DLPFC)", "external"): g100(he_d, R["dlpfc"]),
               ("RNA<->text(synthetic) (DLPFC)", "external"): g100(R["dlpfc"], Tx["dlpfc_s"]),
               ("RNA<->mIF(MIPHEI) (DLPFC)", "external"): g100(R["dlpfc"], Zm["dlpfc_v"])}
        for (pair, kind), r in res.items():
            rows.append({"config": cname, "fold": k, "pair": pair, "kind": kind, **r})
        print(f"fold {k} {cname:<10} " + "  ".join(f"{p.split(' (')[0]} {r['R@1']:.3f}" for (p, _), r in res.items()), flush=True)
    pd.DataFrame(rows).to_csv(os.path.join(ROOT, "results", "final_ens_long.csv"), index=False)

R = pd.DataFrame(rows)
S = R.groupby(["config", "kind", "pair"], sort=False)[["R@1", "R@10", "pair_acc"]].agg(["mean", "std"]).round(4)
S.to_csv(os.path.join(ROOT, "results", "final_ens_summary.csv"))
tests = []
for c in CONF:
    if c == "S4":
        continue
    for p in R.pair.unique():
        a = R[(R.config == c) & (R.pair == p)].sort_values("fold")["R@1"].values
        b = R[(R.config == "S4") & (R.pair == p)].sort_values("fold")["R@1"].values
        tests.append({"config": c, "pair": p, "S4": b.mean(), "new": a.mean(), "delta": (a - b).mean(),
                      "folds_won": int((a > b).sum()), "p_paired_t": stats.ttest_rel(a, b).pvalue})
pd.DataFrame(tests).round(4).to_csv(os.path.join(ROOT, "results", "final_ens_tests.csv"), index=False)
pd.set_option("display.width", 250)
print(R.pivot_table(index="config", columns="pair", values="R@1", sort=False).round(3).T.to_string())
print(pd.DataFrame(tests).round(4).to_string())