Blame for experiments/44_rnafm_eval.py

"""RNA encoder comparison + fusion under the final protocol (real pairs only; virtual data only at test time).

RNA encoders (frozen inputs, cache/rnafm/<enc>_<set>.npy from 43_*.py; "hvg" = log1p CP10k of the 2000 HEST HVGs,
identical to 42_train_regimes.py):
  hvg, scgpt, geneformer, nicheformer, tahoex1, aidocell, se600m, c2s, c2s-last, genept-bge, genept-bmc (whatever exists)
  cat:<a>+<b>+...  early fusion: concatenation of the z-scored inputs, each block scaled by 1/sqrt(its dim)
  ens:<a>+<b>+...  late fusion: mean of the L2-normalised outputs of the separately trained single-encoder heads
                   (possible because every head maps into the same frozen UNI2 space)
Head: linear or MLP (42_'s mlp) from the encoder input to the frozen UNI2 space, InfoNCE, AdamW 3e-4 / wd 0.05 /
OneCycle, batch 256, 2700 steps, early stopping on val R@1 every 50 steps (= 42_ S1/S4 RNA recipe).
FM inputs are z-scored with the HEST TRAINING-spot mean / std of fold k (applied unchanged to virtual / external sets).
Per fold k (splits/final/hest_fold<k>.csv, ORION splits/patient_5fold, TCGA splits/final/tcga_fold<k>.csv):
  trained   H&E<->RNA            HEST test spots (real)
  emergent  mIF<->RNA(virtual)   ORION test tiles, mIF = S4 mIF MLP head of fold k (trained here once, cached)
  emergent  text<->RNA(virtual)  TCGA test images, text = S4 linear text head of fold k (clean real text)
  calib     H&E<->RNA(virtual)   HEST test spots with STPath RNA instead of the real one (circular, reference only)
--dlpfc: additionally train on HEST WITHOUT the spatialLIBD study (val = 10% of groups) and test H&E<->RNA on the
         external DLPFC set (482 spots; FM inputs z-scored per dataset as in 38_), mixed galleries G=100.
Usage: python 44_rnafm_eval.py [--encs ...] [--heads linear mlp] [--folds 0 1 2 3 4] [--dlpfc]
Output results/rnafm_long.csv (resumable), results/rnafm_summary.csv
"""
import argparse
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.optimize import linear_sum_assignment

from common import CACHE, ROOT, SPLITS

ap = argparse.ArgumentParser()
ap.add_argument("--encs", nargs="+", default=None)
ap.add_argument("--heads", nargs="+", default=["linear", "mlp"])
ap.add_argument("--folds", nargs="+", type=int, default=[0, 1, 2, 3, 4])
ap.add_argument("--dlpfc", action="store_true")
args = ap.parse_args()
dev = "cuda"
H, RF = os.path.join(CACHE, "hest"), os.path.join(CACHE, "rnafm")
G100, DRAWS, BS, LR, EVAL_EVERY, STEPS = 100, 50, 256, 3e-4, 50, 2700
f32 = lambda a: torch.from_numpy(np.asarray(a, dtype=np.float32))
SETS = ["hest", "hest_v", "orion_v", "tcga_v", "dlpfc"]

# ---------------------------------------------------------------- anchors / other modalities (= 42_)
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)
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)
HE_D = HE_D[:, 0] if HE_D.dim() == 3 else HE_D
TXT = f32(np.load(os.path.join(CACHE, "alltxt_text_biomedclip.npy"))[:, 0]).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():
    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")), "hest_v": virt("hest_val"), "orion_v": virt("orion"),
            "tcga_v": virt("tcga_text"), "dlpfc": d}


def load_enc(e):
    e = e.split("@")[0]  # "<enc>@<n>": same input, head trained with another seed (seed-ensemble control)
    if e == "hvg":
        return {s: f32(x).to(dev) for s, x in hvg_inputs().items()}
    return {s: f32(np.load(os.path.join(RF, f"{e}_{s}.npy"))).to(dev) for s in SETS}


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):
    """modality head X -> frozen anchor space A (rows paired). A_* may be (n, V, 1536) -> random view per sample."""
    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 = F.normalize(A_va[:, 0] if A_va.dim() == 3 else A_va, dim=-1)
    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) * F.normalize(a, dim=-1) @ F.normalize(h(X_tr[b]), dim=-1).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, F.normalize(h(X_va), dim=-1))
            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, best


class Fold:
    def __init__(self, k):
        s = pd.read_csv(os.path.join(SPLITS, "patient_5fold", f"fold{k}.csv"))
        self.o = {x: s.tile_id[s.split == x].values for x in ["train", "val", "test"]}
        mu, sd = IF_O[self.o["train"]].mean((0, 1)), IF_O[self.o["train"]].std((0, 1)) + 1e-3
        self.IF = (IF_O - mu) / sd
        h = pd.read_csv(os.path.join(SPLITS, "final", f"hest_fold{k}.csv"))
        self.h = {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
        self.t = {x: T.index[tsp == x].values for x in ["train", "val", "test"]}
        self.t_va = one_text_per_image(self.t["val"], np.random.default_rng(0))
        self.t_te = one_text_per_image(self.t["test"], np.random.default_rng(k))
        # DLPFC-held-out training pool (= 38_)
        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
        self.d = {"train": pool[~isva], "val": pool[isva]}

    def other_heads(self, k):
        """S4 heads of fold k for mIF (MLP) and text (linear); cached in runs/rnafm/other_fold<k>.pt"""
        f = os.path.join(ROOT, "runs", "rnafm", f"other_fold{k}.pt")
        hm, ht = mlp(512).to(dev), nn.Linear(512, 1536).to(dev)
        if os.path.exists(f):
            sd = torch.load(f, map_location=dev); hm.load_state_dict(sd["mif"]); ht.load_state_dict(sd["text"])
            return hm.eval(), ht.eval()
        o, t = self.o, self.t
        hm, _ = train_head(self.IF[o["train"]].flatten(0, 1), HE_O[o["train"]].flatten(0, 1),
                           self.IF[o["val"], 0], HE_O[o["val"], 0], "mlp", 400, k)
        ii, tt = self.t_va
        ht, _ = train_head(TXT[t["train"]], HE_T[T_IMG[torch.as_tensor(t["train"], device=dev)]],
                           TXT[torch.as_tensor(tt, device=dev)], HE_T[torch.as_tensor(ii, device=dev), 0],
                           "linear", 1000, k, pos_tr=T_IMG[torch.as_tensor(t["train"], device=dev)])
        os.makedirs(os.path.dirname(f), exist_ok=True)
        torch.save({"mif": hm.state_dict(), "text": ht.state_dict()}, f)
        return hm, ht


def prep(X, tr, per_dataset_ext=False):
    """z-score with HEST-train stats (hvg: unchanged, = 42_). Returns dict set -> tensor."""
    if X is None:
        return None
    mu, sd = X["hest"][tr].mean(0), X["hest"][tr].std(0) + 1e-6
    out = {s: (x - mu) / sd for s, x in X.items()}
    if per_dataset_ext:
        out["dlpfc"] = (X["dlpfc"] - X["dlpfc"].mean(0)) / (X["dlpfc"].std(0) + 1e-6)
    return out


def build(enc, fold_tr, dl_tr=None):
    """input tensors per set for encoder spec enc; hvg kept raw (42_), FMs / fusions z-scored."""
    parts = enc.split(":", 1)[1].split("+") if enc.startswith("cat:") else [enc]
    main, ext = [], []
    for p in parts:
        X = CACHE_ENC[p]
        if p.split("@")[0] == "hvg" and len(parts) == 1:
            main.append(X)
            if dl_tr is not None:
                ext.append(prep(X, dl_tr, True))
            continue
        w = 1 / np.sqrt(X["hest"].shape[1]) if len(parts) > 1 else 1.0
        main.append({s: v * w for s, v in prep(X, fold_tr).items()})
        if dl_tr is not None:
            ext.append({s: v * w for s, v in prep(X, dl_tr, True).items()})
    cat = lambda L: {s: torch.cat([d[s] for d in L], 1) for s in SETS}
    return cat(main), (cat(ext) if dl_tr is not None else None)


# ---------------------------------------------------------------- run
avail = ["hvg"] + sorted({f.rsplit("_", 1)[0] for f in os.listdir(RF) if f.endswith("_hest.npy")
                          and all(os.path.exists(os.path.join(RF, f.replace("_hest.npy", f"_{s}.npy"))) for s in SETS)})
encs = sorted(args.encs or avail, key=lambda e: e.startswith("ens:"))  # late fusions after their single heads
singles = sorted({p for e in encs for p in (e.split(":", 1)[1].split("+") if ":" in e else [e])})
CACHE_ENC = {e: load_enc(e) for e in singles}
print("encoders:", {e: CACHE_ENC[e]["hest"].shape[1] for e in singles}, flush=True)
f_out = os.path.join(ROOT, "results", "rnafm_long.csv")
rows = pd.read_csv(f_out).to_dict("records") if os.path.exists(f_out) else []
done = {(r["enc"], r["head"], r["fold"]) for r in rows}
nz = lambda x: F.normalize(x, dim=-1)
for k in args.folds:
    F_ = Fold(k)
    hm, ht = F_.other_heads(k)
    te_o = F_.o["test"]; ii, tt = (torch.as_tensor(x, device=dev) for x in F_.t_te)
    with torch.no_grad():
        z_mif = nz(hm(F_.IF[te_o, 0])); z_txt = nz(ht(TXT[tt]))
    heads_out = {}  # (enc, head) -> dict of output embeddings, for late fusion
    for enc in encs:
        for hk in args.heads:
            parts = enc.split(":", 1)[1].split("+") if enc.startswith("ens:") else None
            need = any(enc in e.split(":", 1)[1].split("+") for e in encs if e.startswith("ens:"))
            if (enc, hk, k) in done and parts is None and not need:
                continue
            tr, va, te = F_.h["train"], F_.h["val"], F_.h["test"]
            if parts is not None:  # late fusion of single-encoder heads trained in this run
                if not all((p, hk) in heads_out for p in parts) or (enc, hk, k) in done:
                    continue
                Z = {s: nz(sum(heads_out[(p, hk)][s] for p in parts)) for s in heads_out[(parts[0], hk)]}
                best = float("nan")
            else:
                X, Xd = build(enc, tr, F_.d["train"] if args.dlpfc else None)
                sd_ = k + 1000 * int(enc.split("@")[1]) if "@" in enc else k
                h, best = train_head(X["hest"][tr], HE_H[tr], X["hest"][va], HE_H[va], hk, STEPS, sd_)
                with torch.no_grad():
                    Z = {"hest": nz(h(X["hest"][te])), "hest_v": nz(h(X["hest_v"][te])),
                         "orion_v": nz(h(X["orion_v"][te_o])), "tcga_v": nz(h(X["tcga_v"][ii]))}
                if args.dlpfc:
                    hd, _ = train_head(Xd["hest"][F_.d["train"]], HE_H[F_.d["train"]], Xd["hest"][F_.d["val"]],
                                       HE_H[F_.d["val"]], hk, STEPS, sd_)
                    with torch.no_grad():
                        Z["dlpfc"] = nz(hd(Xd["dlpfc"]))
                heads_out[(enc, hk)] = Z
            res = {("H&E<->RNA (HEST)", "trained"): g100(nz(HE_H[te]), Z["hest"]),
                   ("mIF<->RNA(virtual) (ORION)", "emergent"): g100(z_mif, Z["orion_v"]),
                   ("text<->RNA(virtual) (TCGA)", "emergent"): g100(z_txt, Z["tcga_v"]),
                   ("H&E<->RNA(virtual) (HEST)", "calib"): g100(nz(HE_H[te]), Z["hest_v"])}
            if "dlpfc" in Z:
                res[("H&E<->RNA (DLPFC external)", "external")] = g100(nz(HE_D), Z["dlpfc"])
            if (enc, hk, k) in done:  # retrained only to feed a late fusion
                continue
            for (pair, kind), r in res.items():
                rows.append({"enc": enc, "head": hk, "fold": k, "pair": pair, "kind": kind, "val_R@1": best, **r})
            pd.DataFrame(rows).to_csv(f_out, index=False)
            print(f"fold {k} {enc:<34} {hk:<6} val {best:.3f}  " +
                  "  ".join(f"{p.split(' (')[0]} {r['R@1']:.3f}" for (p, _), r in res.items()), flush=True)

R = pd.DataFrame(rows)
S = R.groupby(["enc", "head", "kind", "pair"], sort=False)[["R@1", "R@10", "pair_acc"]].agg(["mean", "std"]).round(4)
S.to_csv(os.path.join(ROOT, "results", "rnafm_summary.csv"))
pd.set_option("display.width", 250)
print(R.pivot_table(index=["enc", "head"], columns="pair", values="R@1", sort=False).round(3).to_string())