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

节点 n5

mix 后加两遍阻尼 NN 坐标位移：全局 kNN(α=0.3) 收缩两朵云 + 同型 kNN(α2=3.0,k=10) 把同类型跨阶段细胞共位化，表达不动，位移后 RMS 偏离目标 >5% 才重缩放。

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 599d1d8e9a: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：父节点版本 80df7fc802。改动的文件：solution/EXECUTION.json +1 −0、solution/METHOD.md +44 −0、solution/run.py +109 −10

diff --git a/solution/EXECUTION.json b/solution/EXECUTION.jsonnew file mode 100644index 0000000..9d5125c--- /dev/null+++ b/solution/EXECUTION.json@@ -0,0 +1 @@+{"gpu": false}diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..34cb3ea--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,44 @@+mix 后加两遍阻尼 NN 坐标位移：全局 kNN(α=0.3) 收缩两朵云 + 同型 kNN(α2=3.0,k=10) 把同类型跨阶段细胞共位化，表达不动，位移后 RMS 偏离目标 >5% 才重缩放。++## 方法（family T2HI-02，PLAN 的 NN 位移，加了同型配对分支）++1. 母节点 mix 管线原样保留：procrustes（共有类型质心 Kabsch，z 保持切片轴只定号）→ 两朵云缩放到 log 线性目标 RMS → 按 (1−t, t) 分型分层抽真实细胞，表达和坐标一起走。+2. 新增位移（`nn_displace`，scipy cKDTree）：+   - 第一遍（全局）：每个 A 细胞向选中 B 细胞的 k=10 近邻质心位移 α·t·(nn_mean−pos)，B 细胞对称位移 α·(1−t)，α=0.3；+   - 第二遍（同型）：在位移后的云上，每个细胞只向**同 celltype** 的对侧细胞 kNN 质心位移，α2=3.0；对侧没有该类型（或该型细胞数为 0）的细胞不动。同型质心用逐类型 cKDTree 计算，标签由复现 `mix_indices(seed)` 得到（与 interpolate 内部选择逐位一致，有 assert 保护）。+   - 表达完全不动，每个细胞保留真实测量向量。+3. 位移后 |log(RMS_out/RMS_target)|>0.05 时统一重缩放回目标 RMS（本次运行 RMS 346.3→331.3，log 差 0.044，未触发重缩放）。+4. 单输入退路：b 为 None 时完全跳过位移，与母节点相同。+5. 确定性：cKDTree 查询确定，rng 只来自 --seed；seed 7 复跑逐位一致（已验证）。++## 机制生效证据（α=0.3/α2=3.0/k=10，proxy seed 0）++- 位移非零且空间异质：A 细胞平均位移 33.7、B 细胞 86.4 个坐标单位（云 RMS 346），是同型跨阶段距离驱动的逐细胞向量，不是每型一个常向量（同型第二遍内每个细胞的位移取决于自己的 kNN 几何）。+- RMS 346.3→331.3，向真值方向（proxy 真值 E8.75 RMS≈217，log 比 0.47→0.41）移动。+- neighborhood_mmd raw 0.08286→0.0789（本节点所有配置中同型强位移一致改善：α2=2.0 时 0.0784）；local_spatial 组 54.03→54.85。+- d2_shape 0.0401→0.0369、occupancy_dice 0.834→0.841、shape_scale 组 53.37→54.67。+- 表达四项（de_score/de_direction/mmd_u/variogram）raw 与母节点一致（组内差异仅评分器抽样噪声），证实表达未被触碰。++## 机制关闭对照（mechanism_off_control）++`T2_NN_ALPHA=0` → 位移向量全零，输出与母节点预测**逐位相同**（`np.array_equal` 验证过 X 与坐标），即对照分 = 母节点 59.50。差 0 分 < 1 分，管线无 bug。开关为环境变量，提交默认打开（α=0.3, α2=3.0）。++## vec-score 网格（proxy，A 半，seed 0；噪声约 ±1）++| 配置 | 榜分 | local_spatial | shape_scale | nbd raw |+|---|---:|---:|---:|---:|+| 母节点 mix（=α=0 对照） | 59.50 | 54.03 | 53.37 | 0.08286 |+| 全局 α=0.1 / 0.3 | 59.14 / 59.35 | 53.17 / 52.55 | 53.65 / 55.10 | 0.0844 / 0.0865 |+| 同型 α=0.3 / 0.6 / 1.0 (k=5) | 59.18 / 59.22 / 59.00 | 53.7 | ~53.3 | 0.0826 |+| 全局0.3+同型0.3 (k=10) | 59.33 | 52.81 | 54.75 | 0.0856 |+| 全局0.3+同型1.0 (k=10) | 59.42 | 53.50 | 54.42 | 0.0833 |+| 全局0.3+同型2.0 (k=10) | 59.51 | 55.01 | 53.28 | 0.0784 |+| **全局0.3+同型3.0 (k=10)** | **59.82** | 54.85 | 54.67 | 0.0789 |+| 全局0.3+同型4.0 (k=10) | 58.30 | 52.86 | 50.58 | 0.0854 |++## 验证过 / 没验证++- 验证过：α=0 对照逐位复现母节点；seed 确定性（seed 0/7）；vec-check 通过；上表 9 个查分。+- 没验证：k 在最优 α2 下的敏感性（只测了 k=5/10）；多 seed 稳定性（59.82 比 59.50 高 0.32，在 ±1 噪声内，但 local_spatial +0.8、shape +1.3 两组同向改善，且 nbd raw 改善方向在同型家族里一致）；final 视图（E8.25+E8.75→E8.5，31 个共有类型，同型配对会更可靠、位移场更平滑）。+- 风险：α2=3.0 时 B 细胞平均位移 86（云 RMS 的 25%），同型跨阶段错位大时可能把细胞推过对侧；α2=4.0 已崩（58.3），最优点在过冲边缘。final 上类型更多、跨阶段同型距离更小，位移会更温和，方向应一致。+- 生物学依据：同型细胞跨阶段共位化对应"同一谱系/类型区域在发育中保持空间毗邻"的通用组织学知识；未使用任何保留阶段数据、比例或标记基因（来源：无外部文献，仅评分反馈驱动的参数选择）。diff --git a/solution/run.py b/solution/run.pyindex 8c61767..1d71332 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,28 +1,117 @@ #!/usr/bin/env python3-"""mix (T2 interpolation): real cells from both bracketing inputs, drawn (1−t, t).--Brackets the target with the nearest inputs before and after it, puts both in-one frame (``ALIGN``), rescales both clouds to the log-linear RMS-exp(log r_a + SCALE_DAMP·t·Δlog r), and draws cells stratified by type:-round(t·n) from the later stage, the rest from the earlier one. Expression and-coordinates travel together. n is log-linear in t, clipped to the board range.-Parameters are the T2 card's choice for this board (selected_params.json).-If the target is not bracketed, falls back to the latest input before it.+"""mix + damped NN coordinate displacement (T2 interpolation).++Same as the parent mix: brackets the target with the nearest inputs before and+after it, aligns both clouds in one frame (``ALIGN`` = procrustes on shared-type+centroids, z held), rescales to the log-linear RMS, draws cells stratified by+type ((1−t, t)), expression and coordinates travel together.++New step after sampling (two damped NN-displacement passes on coordinates):+pass 1 (global, alpha=0.3): each stage-A cell is displaced by alpha·t·(kNN+centroid in stage B − pos), stage-B cells symmetrically by alpha·(1−t). Pass 2+(same-type, alpha2=3.0): same displacement but the kNN centroid is taken only+over other-stage cells of the SAME celltype (per-type cKDTree; cells whose type+is absent on the other side do not move). k=10. Expression is NOT touched —+every cell keeps its real measured profile. After displacement, coordinates are+rescaled to the log-linear target RMS only when |log(RMS_out/RMS_target)|>0.05,+keeping scale_log_ratio neutral.++Mechanism off-control: set env T2_NN_ALPHA=0 — all displacement vectors become+zero and the pipeline is bit-identical to the parent mix (verified). Overrides:+T2_NN_ALPHA2, T2_NN_K, T2_NN_SAME_TYPE (0=global only, 1=same-type only,+2=both, default 2). Single-input fallback (b is None): no displacement. """  from __future__ import annotations  import argparse import json+import os import sys  import numpy as np +from src.task2_spatial.frame import rms_radius, scale_to_rms from src.task2_spatial.methods import interpolate from src.task2_spatial.sample import take from src.task2_spatial.view_io import board_params, interp_bracket, load_manifest, panel_genes, read_stage, write_t2  PARAMS = {"align": "procrustes", "scale_damp": 1.0}+ALPHA = float(os.environ.get("T2_NN_ALPHA", "0.3"))+K = int(os.environ.get("T2_NN_K", "10"))+ALPHA2 = float(os.environ.get("T2_NN_ALPHA2", "3.0"))+++SAME_TYPE = bool(int(os.environ.get("T2_NN_SAME_TYPE", "2")))+++def _nn_mean(src_pts, src_lab, query_pts, query_lab, k, same_type):+    """k-NN centroid of each query point in src; per-type trees when same_type."""+    from scipy.spatial import cKDTree++    if not same_type:+        kk = max(1, min(k, src_pts.shape[0]))+        idx = np.atleast_1d(cKDTree(src_pts).query(query_pts, k=kk)[1]).reshape(query_pts.shape[0], kk)+        return src_pts[idx].mean(axis=1)+    out = np.array(query_pts, dtype=np.float64)+    src_lab = np.asarray(src_lab).astype(str)+    query_lab = np.asarray(query_lab).astype(str)+    trees = {}+    for lab in np.unique(src_lab):+        m = np.flatnonzero(src_lab == lab)+        if m.size:+            trees[lab] = (m, cKDTree(src_pts[m]))+    for lab in np.unique(query_lab):+        if lab not in trees:+            continue+        qm = np.flatnonzero(query_lab == lab)+        m, tree = trees[lab]+        kk = max(1, min(k, m.size))+        idx = np.atleast_1d(tree.query(query_pts[qm], k=kk)[1]).reshape(qm.size, kk)+        out[qm] = src_pts[m][idx].mean(axis=1)+    return out+++def nn_displace(coords, n_a, t, alpha, k, target_rms, labels=None):+    """Displace each cell a damped fraction toward its k-NN centroid in the other stage.++    SAME_TYPE=1: per-type k-NN centroids; SAME_TYPE=2: global first, then a+    second pass with per-type k-NN centroids on the displaced cloud.+    """+    ca = np.asarray(coords[:n_a], dtype=np.float64)+    cb = np.asarray(coords[n_a:], dtype=np.float64)+    out = np.array(coords, dtype=np.float64)+    disp_a_norm = disp_b_norm = 0.0+    rms_pre = rms_radius(out)+    lab_a = labels[:n_a] if labels is not None else None+    lab_b = labels[n_a:] if labels is not None else None+    mode = int(os.environ.get("T2_NN_SAME_TYPE", "2"))+    if alpha > 0.0 and ca.shape[0] > 0 and cb.shape[0] > 0:+        passes = {0: [(False, alpha)], 1: [(True, alpha)], 2: [(False, alpha), (True, ALPHA2)]}[mode]+        for same_type, al in passes:+            a2 = out[:n_a]+            b2 = out[n_a:]+            m2 = _nn_mean(b2, lab_b, a2, lab_a, k, same_type)+            out[:n_a] = a2 + al * t * (m2 - a2)+            m2 = _nn_mean(a2, lab_a, b2, lab_b, k, same_type)+            out[n_a:] = b2 + al * (1.0 - t) * (m2 - b2)+        disp_a_norm = float(np.linalg.norm(out[:n_a] - ca, axis=1).mean())+        disp_b_norm = float(np.linalg.norm(out[n_a:] - cb, axis=1).mean())+    rms_post = rms_radius(out)+    rescaled = abs(np.log(max(rms_post, 1e-8) / max(target_rms, 1e-8))) > 0.05+    if rescaled:+        out = scale_to_rms(out, target_rms)+    info = {+        "nn_alpha": alpha,+        "nn_k": k,+        "disp_a_mean": disp_a_norm,+        "disp_b_mean": disp_b_norm,+        "rms_pre_disp": rms_pre,+        "rms_post_disp": rms_post,+        "rms_rescaled": bool(rescaled),+        "rms_final": rms_radius(out),+    }+    return out, info   def main() -> None:@@ -45,8 +134,18 @@ def main() -> None:     stage_b = read_stage(args.data, b, genes)     params = board_params(manifest, "mix", PARAMS, args.seed)     expr, coords, info = interpolate(stage_a, stage_b, t, params)+    n_a = int(info["n_from_a"])+    sel_labels = None+    if SAME_TYPE:+        from src.task2_spatial.sample import mix_indices++        rng = np.random.default_rng(int(params["seed"]))+        ia, ib = mix_indices(stage_a.labels, stage_b.labels, t, int(info["n"]), rng)+        assert ia.size == n_a, "selection replay mismatch"+        sel_labels = np.concatenate([np.asarray(stage_a.labels)[ia], np.asarray(stage_b.labels)[ib]])+    coords, disp_info = nn_displace(coords, n_a, t, ALPHA, K, float(info["target_rms"]), sel_labels)     keep = {k: info.get(k) for k in ("t", "n", "rms_a", "rms_b", "out_rms", "n_shared_types", "z_dot", "z_flipped", "align")}-    print(json.dumps({"bracket": [a["stage"], b["stage"]], **keep}, default=float), file=sys.stderr)+    print(json.dumps({"bracket": [a["stage"], b["stage"]], **keep, **disp_info}, default=float), file=sys.stderr)     write_t2(args.out, expr, coords, genes, seed=args.seed)  

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**：在父节点 mix 管线上按 PLAN 加了 kNN 坐标位移（表达不动），扩展为两遍：全局 kNN 收缩（α=0.3）+ 同型跨阶段 kNN 共位化（α2=3.0，k=10，逐类型 cKDTree，标签由复现 mix_indices 得到）。α2 网格 {0.3→4.0}、模式 {全局/同型/两遍}，共 9 次查分。

**查分结果**：最优（全局0.3+同型3.0, k=10）= **59.82**（母 59.50）；local_spatial 54.03→54.85（nbd raw 0.08286→0.0789），shape_scale 53.37→54.67，表达四项 raw 不变。对照 T2_NN_ALPHA=0 与母节点输出逐位相同。α2=4.0 过冲崩到 58.30。seed 确定、vec-check 通过、EXECUTION.json gpu=false。

**下一步最值得试**：(1) 在最优 α2 附近细扫（2.5/3.5）+ 多 seed 确认 59.8 不是噪声；(2) 同型位移的过冲保护——把位移截断到跨阶段同型质心距（B 细胞平均位移已达云 RMS 的 25%）；(3) 同型配对换成类型内 OT/软配对，final 上 31 个共有类型时该机制应更平滑。

调研员

 replace NN as the matching step. 3) Displacement changes pairwise distances, potentially hurting d2_shape if the target shape is not between the two stages — check d2_shape raw alongside neighborhood. 4) Score noise ~1 point: any improvement < 2 points on a single query is unreliable; re-query best config twice.",
  "family_id": "T2HI-02",
  "mechanism": "After mix sampling, displace each selected cell's coordinates a damped fraction toward its nearest-neighbour centroid in the other stage, using real (un-interpolated) expression. This shifts the cloud toward intermediate spatial positions while preserving each cell's real expression profile, improving expression–position pairing and overall shape.",
  "vs_constant_shift": "A per-type constant shift moves every cell of a type by the same vector, preserving all within-type relative geometry. Here each cell gets an individual displacement proportional to its local cross-stage spatial correspondence (distance and direction to its NN in the other stage). The displacement field is spatially varying and cell-specific, not a single vector per type.",
  "mechanism_evidence": "1) Mean and std of displacement vectors should be > 0 and spatially heterogeneous (not a single constant vector); report per-cell displacement histogram. 2) Output RMS radius should move toward the truth's RMS compared to the undisplaced mix. 3) neighborhood_mmd raw value should decrease (improve) relative to parent's 0.08286. 4) d2_shape and occupancy_dice raw values should also improve. 5) Expression metrics (de_score, de_direction, mmd_u, variogram) should remain unchanged (±noise), confirming expression was untouched.",
  "mechanism_off_control": "Set alpha = 0. All displacement vectors become zero; the pipeline is identical to the parent mix method. Score should reproduce node 2 within ±1 point. Any difference > 1 point indicates a bug in the plumbing (alignment, sampling, or output) rather than the displacement mechanism.",
  "sources": []
}
```

审查员

REVIEW.json written