总览 · ← 返回运行 20261003-094241-search-t2-heart-interp-g24-D-s2

节点 n14

T2HI-02：组成配额混抽基座 + OT耦合重心投影的kNN平滑坐标位移场（w=0.3，只动早期细胞坐标，表达原样）。

方法说明?节点程序自带的 METHOD.md：这个程序做了什么、为什么。

来自 programs.git a4764f8d05:solution/METHOD.md

调研员的计划

代码改动?这个节点的程序和父节点程序的逐行差别：绿色是新增，红色是删除。

对比：这个提交的上一版（种子程序：相对空仓库）。改动的文件：solution/EXECUTION.json +1 −0、solution/METHOD.md +32 −0、solution/run.py +270 −0

diff --git a/solution/EXECUTION.json b/solution/EXECUTION.jsonnew file mode 100644index 0000000..6d8012e--- /dev/null+++ b/solution/EXECUTION.json@@ -0,0 +1 @@+{"gpu": false}\ No newline at end of filediff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..fe309c4--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,32 @@+T2HI-02：组成配额混抽基座 + OT耦合重心投影的kNN平滑坐标位移场（w=0.3，只动早期细胞坐标，表达原样）。++## 方法++1. **基座（= 关闭机制的对照，T2_OT_DISP=0）**：`interp_bracket` 取目标两侧输入（本视图 E8.25 + E9.5，t=0.4）；`procrustes` 对齐（保 z，仅 xy 旋转/反射 + z 符号，共有类型仅 5 个）；细胞类型组成在 log 份额空间（多类 logit = log p，几何插值）插值定配额，含伪计数 0.5/n；每类型的早期/晚期来源数用带容量约束的比例拟合对齐全局 (1−t, t) 拆分；按配额无放回抽真实细胞，表达与坐标原样携带；两朵云各自缩放到 log 线性目标 RMS；输出细胞数 log 线性后夹到 [min_cells, max_cells]。单输入/无括号退路 = 锚定阶段分层照抄。+2. **机制（T2HI-02，默认打开）**：两侧输出细胞各分层抽 ≤5000 锚点，在表达 PCA(30 维, eigh 确定解) 上跑不平衡 Sinkhorn（POT，eps=0.05，reg_m=0.9，代价归一化到均值 1）；早期锚点的目标位置 = 耦合行的晚期坐标重心，位移 d = (重心 − 源位置) × t × w；用坐标 kNN(k=15) 把锚点场平滑广播到全部早期输出细胞；晚期细胞不动；位移后 RMS 仅记录不回缩（回缩会抵消位移带来的收缩收益）。w 默认 0.3（实测最优的打开档位）。++## 机制生效证据++- 位移中位数 43.0（w=0.3）/ 143.5（w=1.0），为坐标 RMS 的 12% / 41%（>2% 阈值）；早期细胞整体向晚期云收缩，混合云 RMS 334.7 → 322.1（w=0.3）→ 301.8（w=1.0）。+- 四组分变化（A 半，seed 0）：表达两组逐位不变（de_score 0.4891、de_direction 0.6479，只动坐标，符合预期）；cell_state 71.59 不变；local_spatial 66.13 → 65.93~66.07（基本持平，kNN 平滑确实保住了邻域，对照方法卡里逐细胞 OT 插值邻域暴跌到 23–41）；shape_scale 50.92 → 49.18（w=0.3）→ 45.50（w=1.0）。+- shape_scale 内部：scale_log_ratio raw 0.433 → 0.394（w=0.3）→ 0.329（w=1.0），收缩方向正确（真值更致密）；但 d2_shape raw 0.046 → 0.057 → 0.106、occupancy_dice 0.812 → 0.798 → 0.793，距离分布被位移扭曲，净效应为负。++## 对照与查分结果（A 半，seed 0，共 9 次查分）++| 配置 | 榜分 | shape_scale | local_spatial |+|---|---:|---:|---:|+| 基座（机制关） | **64.76** | 50.92 | 66.13 |+| OT w=0.3（默认提交） | 64.28 | 49.18 | 65.93 |+| OT w=0.5 / k=30 / eps=0.02+tau=5 | 63.88 / 63.86 / 63.93 | 47.4–47.7 | 66.0 |+| OT w=1.0 | 63.37 | 45.50 | 65.98 |++w 从 0 → 1 单调变差：scale 的收益抵不过 d2_shape/occupancy 的损失。**如实说明：机制按 PLAN 实现且确实改变了细胞坐标（见上），但在本代理上净分数为负（−0.5 ~ −1.4，w=0.3 档在 ~1 分噪声内）**。按 PLAN「提交时保持打开」的要求，默认 w=0.3（实测损失最小的档位）；关闭开关 T2_OT_DISP=0 即纯基座。根因推测：代理括号 E8.25↔E9.5 只有 5 个共有类型，procrustes 对齐差，OT 位移被对齐误差主导；真实括号（31 个共有类型）可能表现不同，但无法在代理上验证。++## 验证过 / 没验证++- 验证：seed 0 两次运行输出逐字节相同（sha256 一致）；vec-check 通过；17616 细胞（= max_cells），稀疏 CSR，坐标有限。+- 未验证：多 seed 的 B 半稳定性；真实括号（E8.25+E8.75→E8.5）上位移的方向是否仍有益（代理的晚期云与真实不同）；eps/k 更细的网格（时间预算内只跑了 PLAN 网格的一角）。++## 知识来源++未使用任何保留阶段/禁窗的测量信息：所有统计量（组成、RMS、对齐、OT 耦合）都现场从视图输入计算；无硬编码细胞数、阶段名或尺寸。生物学知识仅用到「细胞类型标签可用于跨阶段配对/组成插值」这一通用假设（标签来自视图本身）。diff --git a/solution/run.py b/solution/run.pynew file mode 100644index 0000000..a026e84--- /dev/null+++ b/solution/run.py@@ -0,0 +1,270 @@+"""T2 heart interpolation: composition-quota real-cell mix + OT-coupled kNN-smoothed+coordinate displacement field (PLAN family T2HI-02).++Base (mechanism off, env T2_OT_DISP=0):+  - bracket stages (a<t<b) via view_io.interp_bracket; procrustes alignment (z held);+  - per-celltype proportions interpolated in log space (logit/geometric), quotas;+  - per-(stage,type) source counts by IPF against global (1-t, t) stage split;+  - real cells sampled without replacement, expression and coordinates carried as-is;+  - both clouds scaled to log-linear target RMS; output count log-linear, clipped.+Mechanism (on by default):+  - unbalanced Sinkhorn (POT) between expression-PCA(30) embeddings of stratified+    anchor subsamples of the two output sides;+  - barycentric projection of late-stage positions -> per-anchor displacement x t;+  - kNN(k=15)-smoothed broadcast of the anchor field onto every early output cell;+  - late cells held; RMS restored to target if drifted >5%.+Fallbacks: single input / unbracketed target -> stratified copy of anchor stage.+"""++from __future__ import annotations++import argparse+import os+import sys++import numpy as np+from scipy import sparse+from scipy.spatial import cKDTree++from src.task1_temporal.view_io import write_prediction+from src.task2_spatial import view_io+from src.task2_spatial.frame import align_pair, log_interp, rms_radius, scale_to_rms+from src.task2_spatial.transport import stratified_choice++EPS = float(os.environ.get("T2_OT_EPS", "0.05"))+TAU = float(os.environ.get("T2_OT_TAU", "0.9"))+W_DISP = float(os.environ.get("T2_OT_W", "0.3"))+K_SMOOTH = int(os.environ.get("T2_OT_K", "15"))+OT_ANCHORS = 5000+PCA_DIM = 30+SCALE_DAMP = 1.0+++def log(*a):+    print(*a, file=sys.stderr, flush=True)+++def quota_counts(fracs: np.ndarray, n: int, pool_total: np.ndarray) -> np.ndarray:+    """Integer quotas summing to n, capped by available pool per type."""+    raw = fracs * n+    counts = np.floor(raw).astype(np.int64)+    counts = np.minimum(counts, pool_total)+    deficit = int(n - counts.sum())+    if deficit > 0:+        spare = pool_total - counts+        order = np.argsort(-(raw - np.floor(raw)))+        guard = 0+        while deficit > 0 and spare.sum() > 0 and guard < 4 * len(counts) + 10:+            progressed = False+            for i in order:+                if deficit <= 0:+                    break+                if spare[i] > 0:+                    counts[i] += 1+                    spare[i] -= 1+                    deficit -= 1+                    progressed = True+            guard += 1+            if not progressed:+                break+    return counts+++def stage_split(quotas, pool_a, pool_b, n_b_target):+    """Per-type counts from stage b, targeting a global column sum n_b_target."""+    k = len(quotas)+    nb = np.minimum(np.round(quotas * (n_b_target / max(quotas.sum(), 1))).astype(np.int64),+                    np.minimum(quotas, pool_b))+    na = quotas - nb+    over = np.maximum(na - pool_a, 0)+    if over.sum() > 0:+        room = np.minimum(over, quotas - nb)+        nb += room+        na = quotas - nb+    diff = int(n_b_target - nb.sum())+    if diff > 0:+        spare = np.minimum(quotas - nb, pool_b - nb)+        order = np.argsort(-spare)+        for i in order:+            if diff <= 0:+                break+            take = int(min(diff, spare[i]))+            if take > 0:+                nb[i] += take+                diff -= take+    elif diff < 0:+        order = np.argsort(-nb)+        for i in order:+            if diff >= 0:+                break+            give = int(min(-diff, nb[i], quotas[i] - nb[i] + nb[i]))+            give = int(min(give, nb[i]))+            if give > 0 and (quotas[i] - (nb[i] - give)) <= pool_a[i]:+                nb[i] -= give+                diff += give+    nb = np.clip(nb, 0, quotas)+    return nb, quotas - nb+++def sample_pool(pool_idx, n, rng):+    if n <= 0:+        return np.array([], dtype=np.int64)+    if n >= len(pool_idx):+        return np.asarray(pool_idx, dtype=np.int64)+    return rng.choice(np.asarray(pool_idx), int(n), replace=False)+++def ot_displacement(X_e, X_l, coords_e, coords_l, labels_e, labels_l, t, rng):+    """Unbalanced-Sinkhorn coupling on expression PCA -> smoothed displacement for early cells."""+    n_e, n_l = len(coords_e), len(coords_l)+    ia = np.arange(n_e) if n_e <= OT_ANCHORS else stratified_choice(labels_e, OT_ANCHORS, rng)+    ib = np.arange(n_l) if n_l <= OT_ANCHORS else stratified_choice(labels_l, OT_ANCHORS, rng)++    A = np.asarray(X_e[ia].todense(), dtype=np.float64)+    B = np.asarray(X_l[ib].todense(), dtype=np.float64)+    if A.shape[0] + B.shape[0] < 10:+        return np.zeros((n_e, 3))+    Z = np.vstack([A, B])+    mu = Z.mean(0)+    Zc = Z - mu+    dim = min(PCA_DIM, Zc.shape[1], Zc.shape[0] - 1)+    if dim < 2:+        return np.zeros((n_e, 3))+    # randomized-free deterministic SVD on the small gene dimension+    cov = Zc.T @ Zc+    w, V = np.linalg.eigh(cov)+    V = V[:, ::-1][:, :dim]+    Ea = Zc[: len(A)] @ V+    Eb = Zc[len(A):] @ V++    aa = np.einsum("ij,ij->i", Ea, Ea)[:, None]+    bb = np.einsum("ij,ij->i", Eb, Eb)[None, :]+    M = np.maximum(aa + bb - 2.0 * (Ea @ Eb.T), 0.0)+    m = M.mean()+    if not np.isfinite(m) or m <= 0:+        return np.zeros((n_e, 3))+    M /= m++    import ot++    wa = np.full(len(ia), 1.0 / len(ia))+    wb = np.full(len(ib), 1.0 / len(ib))+    P = ot.sinkhorn_unbalanced(wa, wb, M, reg=EPS, reg_m=TAU)+    P = np.nan_to_num(P, nan=0.0, posinf=0.0, neginf=0.0)+    rs = P.sum(1)+    ok = rs > 1e-12+    if ok.sum() < 10:+        log("OT: degenerate coupling, skipping displacement")+        return np.zeros((n_e, 3))+    bary = (P[ok] @ coords_l[ib]) / rs[ok][:, None]+    d_anchor = (bary - coords_e[ia][ok]) * float(t)++    # kNN-smoothed broadcast of the anchor field onto every early cell+    tree = cKDTree(coords_e[ia][ok])+    _, nbr = tree.query(coords_e, k=min(K_SMOOTH, ok.sum()))+    if nbr.ndim == 1:+        nbr = nbr[:, None]+    d = d_anchor[nbr].mean(axis=1)+    d[~np.isfinite(d).any(1)] = 0.0+    return d * W_DISP+++def copy_last(view, manifest, genes, seed, out):+    entry = view_io.anchor_entry(manifest)+    S = view_io.read_stage(view, entry, genes)+    rng = np.random.default_rng(seed)+    hi = int(manifest["max_cells"])+    idx = np.arange(S.n) if S.n <= hi else stratified_choice(S.labels, hi, rng)+    write_prediction(S.X[idx], genes, out, coords=np.asarray(S.coords[idx], dtype=np.float64), seed=seed)+++def main():+    ap = argparse.ArgumentParser()+    ap.add_argument("--data", required=True)+    ap.add_argument("--out", required=True)+    ap.add_argument("--seed", type=int, required=True)+    args = ap.parse_args()+    view, out, seed = args.data, args.out, args.seed+    rng = np.random.default_rng(seed)++    manifest = view_io.load_manifest(view)+    genes = view_io.panel_genes(view, manifest)+    lo, hi = int(manifest["min_cells"]), int(manifest["max_cells"])++    a_entry, b_entry, t = view_io.interp_bracket(manifest)+    if b_entry is None:+        log("no bracket -> stratified copy of anchor stage")+        copy_last(view, manifest, genes, seed, out)+        return++    A = view_io.read_stage(view, a_entry, genes)+    B = view_io.read_stage(view, b_entry, genes)+    log(f"bracket {a_entry['stage']} + {b_entry['stage']} -> t={t:.3f}; "+        f"n_a={A.n} n_b={B.n}")++    # aligned, centred frames (z held)+    ca, cb, info = align_pair(A.coords, B.coords, A.labels, B.labels, method="procrustes")+    rms_t = log_interp(rms_radius(A.coords), rms_radius(B.coords), t, SCALE_DAMP)+    ca = scale_to_rms(ca, rms_t)+    cb = scale_to_rms(cb, rms_t)+    log(f"align: shared_types={info.get('n_shared_types')} z_flipped={info.get('z_flipped')} "+        f"rms_target={rms_t:.1f}")++    n_out = int(np.clip(round(log_interp(A.n, B.n, t)), lo, hi))++    # logit(log-share) composition interpolation with pseudocount+    types = sorted(set(map(str, A.labels)) | set(map(str, B.labels)))+    la = np.asarray(A.labels).astype(str)+    lb = np.asarray(B.labels).astype(str)+    cnt_a = np.array([np.sum(la == k) for k in types], dtype=np.float64)+    cnt_b = np.array([np.sum(lb == k) for k in types], dtype=np.float64)+    eps = 0.5 / max(A.n, B.n)+    pa = cnt_a / A.n + eps+    pb = cnt_b / B.n + eps+    pa /= pa.sum()+    pb /= pb.sum()+    pt = np.exp((1.0 - t) * np.log(pa) + t * np.log(pb))+    pt /= pt.sum()++    idx_a = {k: np.flatnonzero(la == k) for k in types}+    idx_b = {k: np.flatnonzero(lb == k) for k in types}+    pool_a = np.array([len(idx_a[k]) for k in types])+    pool_b = np.array([len(idx_b[k]) for k in types])+    quotas = quota_counts(pt, n_out, pool_a + pool_b)+    n_b_target = int(round(t * n_out))+    nb, na = stage_split(quotas, pool_a, pool_b, n_b_target)+    log(f"n_out={n_out} from_a={na.sum()} from_b={nb.sum()}")++    picks_a, picks_b = [], []+    for i, k in enumerate(types):+        picks_a.append(sample_pool(idx_a[k], na[i], rng))+        picks_b.append(sample_pool(idx_b[k], nb[i], rng))+    sel_a = np.concatenate(picks_a) if picks_a else np.array([], dtype=np.int64)+    sel_b = np.concatenate(picks_b) if picks_b else np.array([], dtype=np.int64)+    sel_a = np.sort(sel_a)+    sel_b = np.sort(sel_b)++    X_out = sparse.vstack([A.X[sel_a], B.X[sel_b]], format="csr")+    coords_out = np.vstack([ca[sel_a], cb[sel_b]]) if len(sel_a) else cb[sel_b].copy()++    if os.environ.get("T2_OT_DISP", "1") != "0" and len(sel_a) >= 20 and len(sel_b) >= 20:+        d = ot_displacement(A.X[sel_a], B.X[sel_b], coords_out[: len(sel_a)],+                            coords_out[len(sel_a):], np.asarray(A.labels)[sel_a],+                            np.asarray(B.labels)[sel_b], t, rng)+        norms = np.linalg.norm(d, axis=1)+        log(f"OT displacement: median={np.median(norms):.2f} p95={np.percentile(norms, 95):.2f} "+            f"rms={rms_t:.1f} rel_median={np.median(norms) / max(rms_t, 1e-9):.4f}")+        coords_out[: len(sel_a)] += d+        log(f"RMS after displacement: {rms_radius(coords_out):.1f} (target was {rms_t:.1f})")+    else:+        log("mechanism OFF (T2_OT_DISP=0): pure composition-quota mix")++    coords_out = np.asarray(coords_out, dtype=np.float64)+    if not np.isfinite(coords_out).all():+        raise RuntimeError("non-finite coordinates")+    write_prediction(X_out, genes, out, coords=coords_out, seed=seed)+    log(f"wrote {X_out.shape[0]} cells")+++if __name__ == "__main__":+    main()

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

用到的知识库条目

分析结果?分析员写的 ANALYSIS.json：改了什么、各组分数怎么变、假设是否成立、经验和下一步建议。

对话摘要?每个角色和大模型对话的统计：轮数、工具调用、用时、token 数和最后的回答摘录；原始记录只给路径。

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

工程师

**做了什么**：从零实现 T2HI-02——log 份额空间组成配额 + IPF 来源拆分的真实细胞混抽基座（procrustes 保 z 对齐），叠加 OT 耦合（不平衡 Sinkhorn，表达 PCA30）重心投影的 kNN(k=15) 平滑坐标位移场，带 T2_OT_DISP 关闭开关；seed 确定（双跑 sha256 一致）、vec-check 通过、纯 CPU 35s。

**查分（A 半，9/20 次）**：基座 64.76（≈节点10 的 65.25，噪声内）；OT 位移 w=0.3→64.28、w=0.5→63.88、w=1.0→63.37，单调变差——位移收缩云使 scale_log_ratio 改善（0.433→0.329）且 kNN 平滑保住了邻域（66），但 d2_shape/occupancy 被对齐误差主导的位移扭曲。按 PLAN 要求提交时机制保持打开（w=0.3），METHOD.md 如实报告对照更高。

**下一步最值得试**：代理括号只有 5 个共有细胞类型导致 procrustes 对齐差、OT 位移方向失真；改进方向是先用表达空间（而非类型质心）做跨阶段软对齐再算位移，或把位移限制在类型内质心残差上；另一条路是基座本身（cell_state 71.6）叠加节点 8 式的收缩型 NN 位移专攻 shape_scale。

调研员

{
  "name": "OT耦合kNN平滑位移场驱动坐标（保邻域hybrid）",
  "motivation": "当前最佳节点10（rank3 65.16）的shape_scale仅51.08，是四组中最弱；local_spatial 65.80仍有提升空间。方法卡显示逐细胞OT坐标插值把邻域打到23-41、总分最高45，但kNN平滑位移场可保邻域同时改善形状。节点5/8证明只动坐标的位移可改善local_spatial和shape_scale而不影响表达两组。本方案在节点10的组成配额基座上叠加OT驱动的平滑坐标位移，专攻shape_scale和local_spatial。",
  "approach": "步骤：\n1. 基座：procrustes对齐（保z轴，仅允许xy旋转+平移+均匀缩放）+ log线性RMS插值 + logit空间组成配额抽真实细胞（同节点10逻辑，表达/坐标原样）。\n2. OT耦合：对两侧阶段细胞在表达PCA前30维上跑unbalanced Sinkhorn（正则eps=0.05，不平衡权重tau=0.9），得耦合矩阵P（n_a × n_b）。用POT库sinkhorn_unbalanced；若超时退化到scipy linear_sum_assignment取前5000对。\n3. 重心投影：对输出中来自早期阶段的每个细胞i，目标位置 = Σ_j P[i,j]*pos_b[j] / Σ_j P[i,j]；位移d_i = (target_pos_i - source_pos_i) × t。\n4. kNN平滑：对每个待位移细胞，取其输出云内k=15最近邻的位移均值作为最终位移（一次迭代即可）。\n5. 应用：early-stage细胞坐标 += smoothed_d_i；late-stage细胞坐标不动。\n6. 位移后检查：若RMS偏离目标>5%则重缩放回目标RMS。\n关键参数初值：sinkhorn eps=0.05（范围0.01-0.1），tau=0.9（范围0.7-1.0），k_smooth=15（范围10-25），t为插值参数。vec-score快速筛选：先跑eps∈{0.02,0.05,0.1}×k∈{10,15,20}共9组，选邻域+形状组最优再细调。\n单输入退路：仅一个阶段时跳过OT，输出=copy_last。",
  "expected_groups": ["shape_scale", "local_spatial"],
  "risks": "1) Sinkhorn在30分钟内可能跑不完（细胞数>2万时）；应尽早检查耗时，超时用linear_sum_assignment或降采样到5000。2) 位移太小（耦合接近均匀）则无改善；Engineer应打印位移范数分布，若中位数<坐标RMS的1%则机制未生效。3) kNN平滑过度可能抹平真实位移；对照：不光滑直接应用看邻域是否暴跌。4) 形状改善可能<1分噪声；需A半3-seed确认。",
  "family_id": "T2HI-02",
  "mechanism": "OT耦合给出跨阶段细胞对应关系，重心投影将早期细胞坐标推向其对应的晚期位置，kNN平滑保证局部邻域整块移动而非逐细胞散乱插值。",
  "vs_constant_shift": "常数位移对所有同类型细胞施加相同向量；本方案的位移由OT耦合逐细胞决定（取决于该细胞在表达空间中与对侧的匹配），且经kNN平滑后是空间连续的场，不同位置、不同表达状态的细胞获得不同位移。",
  "mechanism_evidence": "1) 打印位移范数分布（中位数、P95），应>坐标RMS的2%才有意义；2) 位移后neighborhood_mmd raw应≤无位移基座（节点10的0.0510）；3) d2_shape raw应下降；4) 表达两组（de_score、de_direction）raw应与基座逐位相同（只动坐标）；5) 位移方向与细胞类型迁移方向的一致性（定性检查）。",
  "mechanism_off_control": "设环境变量T2_OT_DISP=0（或位移权重=0），跳过步骤3-5，输出退化为纯组成配额mix（等同节点10逻辑）。预期差别：shape_scale和local_spatial回落到节点10水平（~51/66），表达两组不变。",
  "sources": []
}