Blame for experiments/48_joint_ens.py

"""Joint training of the multi-encoder (late-fusion) four-modality model: does the J8 two-stage gain of 42_ stack with
the encoder fusion of 47_?  Real pairs only; virtual / synthetic data only at test time (same splits / test items).

Stage 1 (separate, frozen UNI2 anchor; = 47_): one head per encoder - VirTues mIF -> MLP; every RNA / text encoder ->
linear (FM inputs z-scored with train stats, hvg / BiomedCLIP raw).
Stage 2 (joint, = J8 of 42_): all stage-1 heads + a shared zero-initialised residual adapter on UNI2 (H&E = UNI2 +
adapter(UNI2)) fine-tuned together, sum of the three pair losses (H&E-mIF, H&E-RNA, H&E-text), similarity
distillation lambda 1, AdamW lr 1e-4, 1000 steps, joint early stopping on the geometric mean of the three val R@1.
Configurations (per fold, all from the SAME stage-1 heads):
  S4        stage 1 only, RNA hvg, text BiomedCLIP                         (single encoders, separate)
  J8        S4 -> stage 2                                                  (single encoders, joint)
  E         stage 1 only, ens-small (RNA hvg+Tahoe-x1+SE-600M+AIDO.Cell, text KEEP+BiomedCLIP+Qwen3-0.6B)
  EJ-fused  E -> stage 2, loss on the FUSED modality embedding normalise(sum_e normalise(head_e(x_e)))
  EJ-each   E -> stage 2, loss = mean over encoders of each head's own loss (fusion only at test time)
  EJ-frz    EJ-each with the RNA / text heads frozen (only the H&E adapter, the mIF head and the temperatures move)
Output results/joint_ens_long.csv, results/joint_ens_summary.csv, results/joint_ens_tests.csv
"""
import copy
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_E, TXT_E = ["hvg", "tahoex1", "se600m", "aidocell"], ["keep", "biomedclip", "qwen3-0.6b"]
CONF = {"S4": (["hvg"], ["biomedclip"], None), "J8": (["hvg"], ["biomedclip"], "fused"),
        "E": (RNA_E, TXT_E, None), "EJ-fused": (RNA_E, TXT_E, "fused"), "EJ-each": (RNA_E, TXT_E, "each"),
        "EJ-frz": (RNA_E, TXT_E, "frz")}

# ---------------------------------------------------------------- data (= 47_)
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)
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)
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]
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]
    return {"hest": np.load(os.path.join(H, "hvg.npy")), "orion_v": virt("orion"), "tcga_v": virt("tcga_text")}


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"]})
          for e in RNA_E}
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"]} for e in TXT_E}


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_ / 47_
    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 clip_loss(lg, pos=None):
    if pos is None:
        t = torch.arange(len(lg), device=dev)
        return (F.cross_entropy(lg, t) + F.cross_entropy(lg.T, t)) / 2
    tgt = pos / pos.sum(1, keepdim=True)
    return (torch.sum(-tgt * F.log_softmax(lg, 1), 1).mean() + torch.sum(-tgt * F.log_softmax(lg.T, 1), 1).mean()) / 2


def zs(X, ref_set, idx, raw):
    if raw:
        return dict(X)
    mu, sd = X[ref_set][idx].mean(0), X[ref_set][idx].std(0) + 1e-6
    return {s: (v - mu) / sd for s, v in X.items()}


class Joint(nn.Module):
    """heads: 'mif' + 'rna:<enc>' + 'text:<enc>'; logits per head + fused 'rna' / 'text'; residual adapter on UNI2"""

    def __init__(self, heads, logits, adapter):
        super().__init__()
        self.heads = nn.ModuleDict({k.replace(".", "_"): copy.deepcopy(h) for k, h in heads.items()})
        self.keys = {k.replace(".", "_"): k for k in heads}
        self.logit = nn.ParameterDict({k.replace(":", "__").replace(".", "_"): nn.Parameter(v.detach().clone())
                                       for k, v in logits.items()})
        self.ah = None
        if adapter:
            self.ah = mlp(1536)
            nn.init.zeros_(self.ah[-1].weight); nn.init.zeros_(self.ah[-1].bias)

    def h(self, k):
        return self.heads[k.replace(".", "_")]

    def lg(self, k):
        return self.logit[k.replace(":", "__").replace(".", "_")].exp().clamp(max=100)

    def he(self, x):
        return x if self.ah is None else x + self.ah(x)

    def fuse(self, mod, encs, X):
        """X: enc -> input rows; returns normalised fused embedding and the per-encoder embeddings"""
        z = {e: nz(self.h(f"{mod}:{e}")(X[e])) for e in encs}
        return nz(sum(z.values())), z


f_out = os.path.join(ROOT, "results", "joint_ens_long.csv")
rows = pd.read_csv(f_out).to_dict("records") if os.path.exists(f_out) else []
done = {(r["config"], r["fold"]) for r in rows}

for k in range(5):
    if all((c, k) in done for c in CONF):
        continue
    torch.manual_seed(k)
    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"]}
    mu_if, sd_if = IF_O[o["train"]].mean((0, 1)), IF_O[o["train"]].std((0, 1)) + 1e-3
    nif = lambda x: (x - mu_if) / sd_if
    hh = pd.read_csv(os.path.join(SPLITS, "final", f"hest_fold{k}.csv"))
    hs = {x: hh.spot_idx[hh.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)
    Xr = {e: zs(RNA_IN[e], "hest", hs["train"], e == "hvg") for e in RNA_E}
    Xt = {e: zs(TXT_IN[e], "alltxt", trt, e == "biomedclip") for e in TXT_E}
    o_tr = torch.as_tensor(o["train"], device=dev)
    h_tr = torch.as_tensor(hs["train"], device=dev)

    def batch(task, rng):
        """one training batch: (H&E anchor rows, modality inputs (enc -> rows), multi-positive matrix or None)"""
        if task == "mif":
            b = o_tr[torch.as_tensor(rng.choice(len(o_tr), BS, replace=False), device=dev)]
            d = torch.randint(0, 8, (BS,), device=dev)
            return HE_O[b, d], {"mif": nif(IF_O[b, d])}, None
        if task == "rna":
            b = h_tr[torch.as_tensor(rng.choice(len(h_tr), BS, replace=False), device=dev)]
            return HE_H[b], {e: Xr[e]["hest"][b] for e in RNA_E}, None
        b = trt[torch.as_tensor(rng.choice(len(trt), BS, replace=False), device=dev)]
        im = T_IMG[b]
        return (HE_T[im, torch.randint(0, 8, (BS,), device=dev)], {e: Xt[e]["alltxt"][b] for e in TXT_E},
                (im[:, None] == im[None, :]).float())

    VAL = {"mif": (HE_O[o["val"], 0], {"mif": nif(IF_O[o["val"], 0])}),
           "rna": (HE_H[hs["val"]], {e: Xr[e]["hest"][hs["val"]] for e in RNA_E}),
           "text": (HE_T[vi, 0], {e: Xt[e]["alltxt"][vt] for e in TXT_E})}

    # ------------------------------------------------ stage 1: one head per encoder, frozen anchor (= 47_)
    def train_head(task, enc, kind, steps):
        torch.manual_seed(k); rng = np.random.default_rng(k)
        din = VAL[task][1][enc].shape[1]
        h = (mlp(din) if kind == "mlp" else nn.Linear(din, 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(VAL[task][0])
        best, state = -1, None
        for step in range(1, steps + 1):
            h.train()
            a, X, pos = batch(task, rng)
            loss = clip_loss(logit.exp().clamp(max=100) * nz(a) @ nz(h(X[enc])).T, pos)
            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(VAL[task][1][enc])))
                if v > best:
                    best, state = v, ({k_: x.detach().clone() for k_, x in h.state_dict().items()}, logit.detach().clone())
        h.load_state_dict(state[0]); h.eval()
        return h, state[1]

    HEADS, LOGITS = {}, {}
    HEADS["mif"], LOGITS["mif"] = train_head("mif", "mif", "mlp", 400)
    for e in RNA_E:
        HEADS[f"rna:{e}"], LOGITS[f"rna:{e}"] = train_head("rna", e, "linear", 2700)
    for e in TXT_E:
        HEADS[f"text:{e}"], LOGITS[f"text:{e}"] = train_head("text", e, "linear", 1000)
    print(f"fold {k} stage 1 done", flush=True)

    # ------------------------------------------------ stage 2: joint residual fine-tune (= J8 of 42_)
    def val_score(m, re_, te_):
        m.eval()
        with torch.no_grad():
            v = {"mif": r1(nz(m.he(VAL["mif"][0])), nz(m.h("mif")(VAL["mif"][1]["mif"]))),
                 "rna": r1(nz(m.he(VAL["rna"][0])), m.fuse("rna", re_, VAL["rna"][1])[0]),
                 "text": r1(nz(m.he(VAL["text"][0])), m.fuse("text", te_, VAL["text"][1])[0])}
        m.train()
        return float(np.exp(np.mean(np.log([max(x, 1e-4) for x in v.values()])))), v

    def joint(re_, te_, mode, steps=1000, lr=1e-4, distill=1.0):
        keys = ["mif"] + [f"rna:{e}" for e in re_] + [f"text:{e}" for e in te_]
        lg = {kk: LOGITS[kk] for kk in keys}
        lg["rna"] = torch.stack([LOGITS[f"rna:{e}"] for e in re_]).mean()
        lg["text"] = torch.stack([LOGITS[f"text:{e}"] for e in te_]).mean()
        m = Joint({kk: HEADS[kk] for kk in keys}, lg, adapter=True).to(dev)
        torch.manual_seed(k); rng = np.random.default_rng(k)
        if mode == "frz":
            for kk in keys[1:]:
                m.h(kk).requires_grad_(False)
        opt = torch.optim.AdamW([p for p in m.parameters() if p.requires_grad], lr=lr, weight_decay=0.05)
        sch = torch.optim.lr_scheduler.OneCycleLR(opt, max_lr=lr, total_steps=steps, pct_start=0.05)
        best, state, bv = -1, None, None
        m.train()
        for step in range(1, steps + 1):
            loss = 0
            for task in ("mif", "rna", "text"):
                a, X, pos = batch(task, rng)
                za = nz(m.he(a))
                if task == "mif":
                    loss = loss + clip_loss(m.lg("mif") * za @ nz(m.h("mif")(X["mif"])).T, pos)
                else:
                    encs = re_ if task == "rna" else te_
                    X = {e: X[e] for e in encs}
                    zf, ze = m.fuse(task, encs, X)
                    if mode == "fused":
                        loss = loss + clip_loss(m.lg(task) * za @ zf.T, pos)
                    else:
                        loss = loss + sum(clip_loss(m.lg(f"{task}:{e}") * za @ ze[e].T, pos) for e in encs) / len(encs)
                u = nz(a)
                loss = loss + distill * F.mse_loss(za @ za.T, u @ u.T)
            opt.zero_grad(set_to_none=True); loss.backward(); opt.step(); sch.step()
            if step % EVAL_EVERY == 0 or step == steps:
                sc, v = val_score(m, re_, te_)
                if sc > best:
                    best, bv, state = sc, v, {k_: x.detach().clone() for k_, x in m.state_dict().items()}
        m.load_state_dict(state); m.eval()
        return m, bv

    @torch.no_grad()
    def evaluate(m, re_, te_):
        R = lambda st, idx: m.fuse("rna", re_, {e: Xr[e][st][idx] for e in re_})[0]
        Tx = lambda st, idx: m.fuse("text", te_, {e: Xt[e][st][idx] for e in te_})[0]
        zm = nz(m.h("mif")(nif(IF_O[o["test"], 0])))
        zt = Tx("alltxt", tt)
        return {("H&E<->mIF (ORION)", "trained"): g100(nz(m.he(HE_O[o["test"], 0])), zm),
                ("H&E<->RNA (HEST)", "trained"): g100(nz(m.he(HE_H[hs["test"]])), R("hest", hs["test"])),
                ("H&E<->text (TCGA)", "trained"): g100(nz(m.he(HE_T[ii, 0])), zt),
                ("mIF<->RNA(virtual) (ORION)", "emergent"): g100(zm, R("orion_v", o["test"])),
                ("mIF<->text(synthetic) (ORION)", "emergent"): g100(zm, Tx("orion_gentext", o["test"])),
                ("text<->RNA(virtual) (TCGA)", "emergent"): g100(zt, R("tcga_v", ii)),
                ("text<->mIF(MIPHEI) (TCGA)", "emergent"): g100(zt, nz(m.h("mif")(nif(IF_V_T[ii]))))}

    for cname, (re_, te_, mode) in CONF.items():
        if (cname, k) in done:
            continue
        if mode is None:  # separate: stage-1 heads, no adapter
            keys = ["mif"] + [f"rna:{e}" for e in re_] + [f"text:{e}" for e in te_]
            m = Joint({kk: HEADS[kk] for kk in keys}, {kk: LOGITS[kk] for kk in keys}, adapter=False).to(dev).eval()
            val = {}
        else:
            m, val = joint(re_, te_, mode)
        for (pair, kind), r in evaluate(m, re_, te_).items():
            rows.append({"config": cname, "fold": k, "pair": pair, "kind": kind, **r,
                         **{f"val_{t}": v for t, v in val.items()}})
        pd.DataFrame(rows).to_csv(f_out, index=False)
        print(f"fold {k} {cname:<9} " + "  ".join(f"{r['pair'].split(' (')[0]} {r['R@1']:.3f}"
                                                  for r in rows if r["config"] == cname and r["fold"] == k), flush=True)

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", "joint_ens_summary.csv"))
tests = []
for c, ref in [("J8", "S4"), ("E", "S4"), ("EJ-fused", "S4"), ("EJ-each", "S4"), ("EJ-fused", "E"), ("EJ-each", "E"),
               ("EJ-fused", "J8"), ("EJ-each", "EJ-fused"), ("EJ-frz", "S4"),
               ("EJ-frz", "E"), ("EJ-frz", "EJ-each")]:
    for p in R.pair.unique():
        a = R[(R.config == c) & (R.pair == p)].sort_values("fold")["R@1"].values
        b = R[(R.config == ref) & (R.pair == p)].sort_values("fold")["R@1"].values
        tests.append({"config": c, "vs": ref, "pair": p, "ref": 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", "joint_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(R.pivot_table(index="config", columns="pair", values="pair_acc", sort=False).round(3).T.to_string())
print(pd.DataFrame(tests).round(4).to_string())