总览 · ← 返回运行 20261002-204523-search-t2-heart-interp-g24q

节点 n3

mix + 类型质心位移（T2HI-02）：实现并对照后机制在代理上单调有害，提交为关闭状态（=父节点 mix，逐位一致）。

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 c035d13c7e: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 ab33f143fe。改动的文件：solution/METHOD.md +45 −0、solution/displace.py +107 −0、solution/run.py +14 −2

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..0427f0d--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,45 @@+mix + 类型质心位移（T2HI-02）：实现并对照后机制在代理上单调有害，提交为关闭状态（=父节点 mix，逐位一致）。++## 做了什么++按 PLAN 在父节点 mix 之上实现类型质心位移插值（`solution/displace.py`）：++1. Procrustes 对齐 + log 线性 RMS 缩放后（与父节点同一帧），对两阶段共有细胞类型分别算质心 c_a、c_b；+2. 时间插值 c_t = (1−t)·c_a + t·c_b；+3. 每个被抽中的细胞按其来源阶段平移 (c_t − c_source)，使同一类型的两个子群落到共同的中间质心，类型内相对几何（邻域）保持刚性不动；+4. 之后照常去重抖动、整体缩回目标 RMS。RNG 消耗顺序与父节点完全一致。++保护条款（PLAN risks）：类型在任一阶段 <10 个细胞不参与；可用共有类型 <3 时整体跳过位移（常数位移会被末端居中消掉，等价于父节点）。开关：环境变量 `T2_DISPLACE=1` 打开、`T2_DISP_GAIN`（默认 1，可负）、`T2_DISP_AXES=xy`（z 为切片轴，可排除）。**提交默认 OFF。**++## 机制生效的证据++proxy（E8.25_late + E9.5 → E8.75，t=0.4，5 个共有类型）上 gain=1 时：逐类型位移范数 mean=151.1、**std=75.4**（各类型位移明显不同，不是常数位移）；确实改变了坐标（`.X` 逐位不变，`obsm` 变）。++## 对照结果（vec-score，A 半，seed 0，同一预测管线）++| 配置 | 榜分 | 表达 | 状态 | 形状 | 邻域 | neighborhood_mmd |+|---|---:|---:|---:|---:|---:|---:|+| DISPLACE=0（对照=父节点） | **59.07** | 63.83 | 65.93 | 53.07 | 53.46 | 0.0834 |+| gain=1 | 58.50 | 63.83 | 65.93 | 53.19 | 51.07 | 0.0917 |+| gain=0.5 | 58.66 | 63.83 | 65.93 | 52.63 | 52.24 | 0.0876 |+| gain=−0.5（反向） | 58.85 | 63.83 | 65.93 | 52.78 | 52.86 | 0.0854 |+| gain=1，仅 xy | 58.35 | 63.83 | 65.93 | 52.83 | 50.83 | 0.0926 |++表达/状态两组在所有变体上逐位相同（位移只动坐标），变化全部来自形状与邻域。位移幅度越大邻域越差，且**单调**：正向（收拢双曝光）、反向（拉开双曝光）、只动 xy 都劣于对照。PLAN 的接受标准（形状 +2 且邻域 +1，其余不回退 >1）未满足 → 拒绝该机制，提交关闭状态。++结论/解读：真实中间阶段（E8.75）的局部邻域结构更接近"两个端点阶段的叠加"，而不是"每类收到单一中间质心"。把类型子云刚性平移会破坏细胞与周围其他类型细胞的相对关系（neighborhood_mmd 0.083→0.092），说明评分器看的邻域是跨类型的，逐类型刚体位移恰好扰乱它。这与方法卡中 OT/位移家族在心脏榜上全线低于 mix 的记录一致。++## 验证过 / 没验证++- 验证：DISPLACE=0 输出与父节点管线逐位一致（`.X` 与 `obsm` 均相同）；seed 0/1 均通过 `vec-check`；机制打开时逐类型位移范数 std>0；单输入阶段回退路径未动（与父节点相同）。运行 ~2.5 s、峰值内存与父节点相当。+- 没验证：位移在 final 视图（31 个共有类型、t=0.5）上的行为——代理上已单调有害，且代理只有 5 个共有类型、质心噪声更大，final 上不一定同样差，但按接受标准不在本节点冒险；也没试逐类型部分位移（只收拢不重合）与按邻域 MMD 直接优化的非刚性变形。++## 知识来源++无外部生物知识；只用视图内数据（细胞类型标签、坐标、时间差）。z 为切片轴、不随 xy 旋转的处理沿用父节点/方法卡。++## 下一步建议++1. 逐类型"部分收拢"（如把 a 子云向 b 子云移动小步长并同时做各向同性局部收缩）可能不行——gain=−0.5 反向也变差，说明该轴上对照就是局部最优；不要再扫质心位移家族。+2. cell_state/expression_change（65.9/63.8）只由抽样细胞决定，与坐标无关；要动它们必须改表达（类型内表达插值，审查意见的方向），但方法卡 hyb_type_mix=58.78 提示全量类型均值混合略差，值得试的是**低权重**（0.1–0.3）类型均值位移而非 1.0。+3. shape_scale 的大头是 scale_log_ratio=0.467（输出 RMS 346 vs 目标 217），代理括号两端都大、原理上够不到 217；用 scale_damp 硬凑会破坏 final（354→217 括号，damp>1 会过缩到 ~105），不要做。diff --git a/solution/displace.py b/solution/displace.pynew file mode 100644index 0000000..eb3e44c--- /dev/null+++ b/solution/displace.py@@ -0,0 +1,107 @@+"""mix + type-wise centroid displacement (T2HI-02).++Same pipeline as ``src.task2_spatial.methods.interpolate`` with+coord_mode="mix", expr_mode="carry" (the parent), plus one extra step:+for every cell type shared by both bracketing stages, its centroid is+interpolated in time, c_t = (1-t)*c_a + t*c_b, in the aligned, RMS-rescaled+frame, and each sampled cell of that type is shifted by c_t - c_source so+both sub-populations of a type land on the same intermediate centre while+keeping their within-type relative geometry (neighbourhoods) intact.+Types with fewer than MIN_TYPE_CELLS cells in either stage are left alone;+if fewer than 3 usable shared types exist the displacement is skipped+(a global shift would be removed by the final centring anyway).+RNG consumption order is identical to the parent, so DISPLACE off / no+shared types reproduces the parent exactly.+"""++from __future__ import annotations++import os++import numpy as np++from src.task2_spatial.frame import align_pair, log_interp, rms_radius, scale_to_rms+from src.task2_spatial.methods import _jitter, _limits, _pack_expr+from src.task2_spatial.sample import mix_indices+from src.task2_spatial.transport import as_dense++MIN_TYPE_CELLS = 10+MIN_SHARED_TYPES = 3+DISP_GAIN = float(os.environ.get("T2_DISP_GAIN", "1.0"))+DISP_AXES = os.environ.get("T2_DISP_AXES", "xyz")+++def _type_centroids(coords: np.ndarray, labels, min_count: int) -> dict:+    labs = np.asarray(labels).astype(str)+    out = {}+    for lab in np.unique(labs):+        idx = np.flatnonzero(labs == lab)+        if idx.size >= min_count:+            out[lab] = coords[idx].mean(axis=0)+    return out+++def interpolate_displace(stage_a, stage_b, t: float, params: dict):+    t = float(t)+    damp = float(params.get("scale_damp", 1.0))+    align = str(params.get("align", "procrustes"))+    rng = np.random.default_rng(int(params.get("seed", 0)))++    aligned_a, aligned_b, info = align_pair(+        stage_a.coords, stage_b.coords, stage_a.labels, stage_b.labels, align+    )+    rms_a = rms_radius(stage_a.coords)+    rms_b = rms_radius(stage_b.coords)+    target_rms = log_interp(rms_a, rms_b, t, damp)+    ca = scale_to_rms(aligned_a, target_rms)+    cb = scale_to_rms(aligned_b, target_rms)+    n = _limits(params, stage_a.n, stage_b.n, t, "interp")+    ia, ib = mix_indices(stage_a.labels, stage_b.labels, t, n, rng)++    cent_a = _type_centroids(ca, stage_a.labels, MIN_TYPE_CELLS)+    cent_b = _type_centroids(cb, stage_b.labels, MIN_TYPE_CELLS)+    shared = sorted(set(cent_a) & set(cent_b))+    active = len(shared) >= MIN_SHARED_TYPES+    la = np.asarray(stage_a.labels).astype(str)+    lb = np.asarray(stage_b.labels).astype(str)++    norms = []+    expr_parts = []+    coord_parts = []+    for idx, src, cs, labs in ((ia, stage_a, ca, la), (ib, stage_b, cb, lb)):+        if not idx.size:+            continue+        expr_parts.append(as_dense(src.X, idx))+        p = np.asarray(cs[idx], dtype=np.float64)+        if active:+            for lab in shared:+                m = labs[idx] == lab+                if not m.any():+                    continue+                ct = (1.0 - t) * cent_a[lab] + t * cent_b[lab]+                d = DISP_GAIN * (ct - cent_a[lab] if src is stage_a else ct - cent_b[lab])+                if DISP_AXES == "xy":+                    d = d.copy()+                    d[2] = 0.0+                p[m] += d+                norms.append(float(np.linalg.norm(d)))+        coord_parts.append(p)++    expr = _pack_expr(expr_parts)+    coords = _jitter(np.vstack(coord_parts), rng)+    coords = scale_to_rms(coords, target_rms)+    info.update(+        t=t,+        n=int(expr.shape[0]),+        rms_a=rms_a,+        rms_b=rms_b,+        target_rms=target_rms,+        out_rms=rms_radius(coords),+        n_from_a=int(ia.size),+        n_from_b=int(ib.size),+        displace=bool(active),+        n_disp_types=len(shared) if active else 0,+        disp_norm_mean=float(np.mean(norms)) if norms else 0.0,+        disp_norm_std=float(np.std(norms)) if norms else 0.0,+    )+    return expr, coords.astype(np.float32), infodiff --git a/solution/run.py b/solution/run.pyindex 8c61767..bee1c75 100644--- a/solution/run.py+++ b/solution/run.py@@ -14,15 +14,18 @@ from __future__ import annotations  import argparse import json+import os import sys  import numpy as np +from displace import interpolate_displace 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}+DISPLACE = os.environ.get("T2_DISPLACE", "0") != "0"   def main() -> None:@@ -44,8 +47,17 @@ def main() -> None:     stage_a = read_stage(args.data, a, genes)     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)-    keep = {k: info.get(k) for k in ("t", "n", "rms_a", "rms_b", "out_rms", "n_shared_types", "z_dot", "z_flipped", "align")}+    if DISPLACE:+        expr, coords, info = interpolate_displace(stage_a, stage_b, t, params)+    else:+        expr, coords, info = interpolate(stage_a, stage_b, t, params)+    keep = {+        k: info.get(k)+        for k in (+            "t", "n", "rms_a", "rms_b", "out_rms", "n_shared_types", "z_dot", "z_flipped", "align",+            "displace", "n_disp_types", "disp_norm_mean", "disp_norm_std",+        )+    }     print(json.dumps({"bracket": [a["stage"], b["stage"]], **keep}, 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

完成。总结：

**做了什么**：按 PLAN（T2HI-02）在父节点 mix 上实现类型质心位移插值（`solution/displace.py`），带 `T2_DISPLACE`/`T2_DISP_GAIN`/`T2_DISP_AXES` 开关，含 PLAN 要求的保护条款（<10 细胞类型不参与、<3 共有类型跳过）。

**查分结果**（A 半，seed 0，共用 5 次额度）：对照（机制关）59.07；gain=1 → 58.50；gain=0.5 → 58.66；gain=−0.5 → 58.85；仅 xy → 58.35。位移逐类型范数 std=75（机制确实生效、非常数位移），但邻域分单调劣化（neighborhood_mmd 0.083→0.092），形状也不涨。PLAN 接受标准未满足 → **提交为机制关闭**，输出与父节点逐位一致（`.X`/`obsm` 均验证相同），seed 0/1 通过 `vec-check`，预期节点分 ≈ 父节点 59.4。METHOD.md 记录完整对照表和负结果。

**下一步最值得试**：(1) 不要再扫质心位移家族——正反方向都变差，对照是局部最优；(2) cell_state/表达组只由抽样细胞决定，试低权重（0.1–0.3）类型均值表达插值（审查意见方向，但避开 hyb_type_mix 的满权重）；(3) shape_scale 受代理括号 RMS 两端大所限，scale_damp 硬凑会破坏 final，不可做。

调研员

ession_change, accept. Total extra compute negligible (<0.1 s).",
  "expected_groups": ["shape_scale", "local_spatial"],
  "risks": "If type centroids are nearly identical across stages (little morphological movement), displacement is ~0 and nothing changes. Engineer should print per-type displacement magnitudes; if all <1e-3 the mechanism is inactive. Also, if the two stages have very few shared types (proxy has only 5), centroid estimates are noisy; Engineer should check n_shared_types and fall back to global shift if <3 types are shared. Risk of slight expression_change regression if coordinate-expression coupling is disrupted, but since expression travels with the cell and only coordinates are shifted, this should not occur.",
  "family_id": "T2HI-02",
  "mechanism": "Type-wise centroid interpolation displaces each cell type's spatial centre to its time-t position between the two bracketing stages, creating a genuinely intermediate shape while preserving within-type local neighbourhoods.",
  "vs_constant_shift": "A constant shift applies one displacement vector to all cells. Here the displacement is type-specific: each cell type has its own (c_t - c_source) vector computed from its own centroids at both stages. Different types move by different amounts and directions, reshaping the cloud rather than translating it rigidly.",
  "mechanism_evidence": "Engineer should log: (a) per-type displacement vector norms — they must differ across types (std > 0); (b) shape_scale and local_spatial sub-scores vs parent; (c) confirm cell_state and expression_change do not drop >1 point. If all displacement norms are identical, the mechanism degenerated to a constant shift.",
  "mechanism_off_control": "Set a flag (e.g. DISPLACE=False) that skips the centroid interpolation step, leaving cells at their source-stage coordinates. This reproduces the parent exactly. Expected: shape_scale and local_spatial return to ~53-54, confirming the displacement was responsible for any gain."
}

审查员

REVIEW.json written