总览 · ← 返回运行 20261003-004922-search-t2-heart-interp-g24

节点 n8

mix 空间分层抽样（T2HI-04）：每共有型 xy k-means 分 bin，bin 内按 (1−t,t) 抽两阶段细胞；local_spatial 未提升（3 种子均值 58.99 vs 父 59.36），机制已生效但方向无效。

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 01d7f77244: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 33a8f2ab38。改动的文件：solution/METHOD.md +38 −0、solution/run.py +25 −10、solution/spatial_mix.py +197 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..2fdb97d--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,38 @@+# mix 空间分层抽样（T2HI-04）：每共有型 xy k-means 分 bin，bin 内按 (1−t,t) 抽两阶段细胞；local_spatial 未提升（3 种子均值 58.99 vs 父 59.36），机制已生效但方向无效。++## 方法++在父节点 2（mix，align=procrustes，scale_damp=1）之上，只改抽样一步：++- 对每个**两阶段共有**的细胞型，把该型全部细胞（两阶段合并，对齐+缩放到目标 RMS 后的坐标）在 xy 上做 k-means，k = max(4, n_type // CELLS_PER_BIN)，CELLS_PER_BIN=100（提交默认）。+- bin 内按该型的 (1−t, t) 配额、以 bin 内各阶段细胞数为权重比例分配（floor + 最大余数 + 容量约束），无放回抽样；表达和坐标随细胞走。+- 非共有型（只出现在一个阶段）与 <20 细胞的型退回父节点的随机抽样；非共有型无法做跨阶段空间配比，这是本机制的天然覆盖上限。+- 单输入阶段（b is None）退路与父节点完全一致。+- 实现：`solution/spatial_mix.py`（`spatial_mix_indices` 复制 `transport.stratified_choice` 的型级配额逻辑，保证每型总数与父节点相同，只改型内选哪些细胞）；`run.py` 里 `SPATIAL_STRATIFY=0` 为关闭机制对照（直接调父节点的 `methods.interpolate`，输出与父节点逐位一致，已验证代码路径相同）。++## 机制生效证据（proxy，E8.25_late+E9.5→E8.75，t=0.4）++- proxy 上共有型只有 5 个（NCC/Peri/V-CM/aPHM/pPHM），占输出 17616 细胞中的约 3553（20%）。+- 源阶段标签的空间自相关：Moran's I 从 0.302（父随机抽样）降到 0.287–0.288（bin100–400）；kNN(k=10) 同源边比例 0.665 → 0.659（随机混合理论值 t²+(1−t)²=0.52，剩余超出部分来自非共有型的纯簇，机制无法触及）。+- 四组分（bin100, seed0）：expression_change 64.07(+0.2)、cell_state 66.17(−0.5)、shape_scale 53.32(−0.05)、local_spatial 53.60(−0.4)，vs 父 63.9/66.7/53.4/54.0。++## 查分记录（A 半，vec-score）++| 配置 | seed0 | seed1 | seed2 | 均值 |+|---|---:|---:|---:|---:|+| 父节点 mix | 59.50 | - | - | 59.36(rank3) |+| bin200 | 58.87 | - | - | - |+| bin400 | 59.17 | - | - | - |+| bin100（提交默认） | 59.29 | 58.74 | 58.93 | 58.99 |+| bin100 + 3D 分箱 | 59.01 | - | - | - |++## 结论与已验证/未验证++- **已验证**：机制按 PLAN 实现并生效（上面 Moran's I / kNN 证据）；bin∈{100,200,400}、2D/3D 分箱都在父节点 ±0.6 分内，无一提升 local_spatial；关闭开关与父节点同路径。+- **结论（负面）**：局部均匀化两阶段混合比不改善 local_spatial，反而 shape_scale 略降（occupancy_dice 0.84→0.81–0.83）。提示真实 E8.75 里"源阶段身份"与空间位置本就相关（发育梯度），强行局部均匀化偏离真实结构；且 proxy 上共有型仅覆盖 20% 细胞，机制杠杆太小。+- **未验证**：final 括号（E8.25+E8.75→E8.5，31 个共有型，覆盖率更高）上机制可能表现不同，本节点无法测。+- 生物学知识来源：仅方法卡与 PLAN 中的坐标系约定（z 为离散切片轴、保 z 的 procrustes），未引入新的先验。++## 下一步建议++放弃"抽样空间均匀化"方向；local_spatial 的瓶颈更可能在非共有型的空间占位（occupancy/variogram），值得试的是对非共有型按目标阶段的空间密度重加权抽样，或直接改坐标生成（如类型级形状插值）而非选细胞。diff --git a/solution/run.py b/solution/run.pyindex 8c61767..7bdd6a7 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,12 +1,16 @@ #!/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).+"""mix + spatial-stratified per-type sampling (T2HI-04).++Same as the parent mix (align=procrustes, scale_damp=1, cells drawn+(1-t, t) stratified by type, expression and coordinates travel together),+except that within each cell type present in both brackets the draw is+spatially stratified: pooled xy of the type is k-means-binned+(CELLS_PER_BIN cells per bin) and every bin contributes stage-a / stage-b+cells in the type's (1-t, t) proportion, so each spatial region of the+output holds the right stage mixture (keeps local neighbourhoods+consistent). Types in one bracket only, or with < 20 cells, keep the+parent's random draw. Set SPATIAL_STRATIFY=0 to fall back to the exact+parent behaviour (mechanism-off control). If the target is not bracketed, falls back to the latest input before it. """ @@ -14,15 +18,23 @@ from __future__ import annotations  import argparse import json+import os import sys  import numpy as np +sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))+ 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 +from spatial_mix import interpolate_spatial+ PARAMS = {"align": "procrustes", "scale_damp": 1.0}+CELLS_PER_BIN = int(os.environ.get("CELLS_PER_BIN", "100"))+BIN_NDIM = int(os.environ.get("BIN_NDIM", "2"))+SPATIAL_STRATIFY = os.environ.get("SPATIAL_STRATIFY", "1") != "0"   def main() -> None:@@ -44,8 +56,11 @@ 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 SPATIAL_STRATIFY:+        expr, coords, info = interpolate_spatial(stage_a, stage_b, t, params, cells_per_bin=CELLS_PER_BIN, ndim=BIN_NDIM)+    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", "n_from_a", "n_from_b")}     print(json.dumps({"bracket": [a["stage"], b["stage"]], **keep}, default=float), file=sys.stderr)     write_t2(args.out, expr, coords, genes, seed=args.seed) diff --git a/solution/spatial_mix.py b/solution/spatial_mix.pynew file mode 100644index 0000000..c4778ed--- /dev/null+++ b/solution/spatial_mix.py@@ -0,0 +1,197 @@+"""mix with spatial-stratified per-type sampling (T2HI-04).++Same pipeline as src.task2_spatial.methods.interpolate for+coord_mode="mix", expr_mode="carry", except that within each cell type+present in BOTH bracketing stages the random draw is replaced by a+spatially stratified draw: pooled xy coordinates of the type are binned+with k-means, and each bin contributes cells from stage a / stage b in+the same proportions as the type-level (1-t, t) allocation, so every+spatial region of the output cloud holds the right stage mixture.+Types seen in only one stage (or too small) keep the parent's random draw.+"""++from __future__ import annotations++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.transport import as_dense++MIN_TYPE_CELLS = 20+++def _alloc_by_type(labels: np.ndarray, n: int):+    """Per-type counts identical to transport.stratified_choice's allocation."""+    labels = np.asarray(labels).astype(str)+    n = int(min(max(n, 1), len(labels)))+    if n >= len(labels):+        return {lab: int(c) for lab, c in zip(*np.unique(labels, return_counts=True))}+    types, counts = np.unique(labels, return_counts=True)+    raw = counts / counts.sum() * n+    alloc = np.floor(raw).astype(int)+    rem = int(n - alloc.sum())+    order = np.argsort(-(raw - alloc))+    for i in range(rem):+        alloc[order[i % len(order)]] += 1+    alloc = np.minimum(alloc, counts)+    deficit = int(n - alloc.sum())+    if deficit > 0:+        spare = counts - alloc+        for i in np.argsort(-spare):+            take = int(min(deficit, spare[i]))+            alloc[i] += take+            deficit -= take+            if deficit == 0:+                break+    return {lab: int(k) for lab, k in zip(types, alloc)}+++def _spread(k: int, weights: np.ndarray, caps: np.ndarray) -> np.ndarray:+    """Split k over bins proportional to weights, capped at caps."""+    out = np.zeros(len(weights), dtype=int)+    total = float(weights.sum())+    if k <= 0 or total <= 0:+        return out+    k = int(min(k, int(caps.sum())))+    raw = weights / total * k+    base = np.minimum(np.floor(raw).astype(int), caps)+    out += base+    rem = k - int(out.sum())+    if rem > 0:+        frac = raw - np.floor(raw)+        spare = caps - out+        order = np.argsort(-frac)+        i = 0+        guard = 0+        while rem > 0 and guard < 10 * len(order) + 10:+            j = order[i % len(order)]+            if spare[j] > 0:+                out[j] += 1+                spare[j] -= 1+                rem -= 1+            i += 1+            guard += 1+    return out+++def _kmeans(xy: np.ndarray, k: int, rng: np.random.Generator, iters: int = 25) -> np.ndarray:+    n = xy.shape[0]+    k = int(min(max(k, 1), n))+    if k <= 1:+        return np.zeros(n, dtype=int)+    cent = xy[rng.choice(n, k, replace=False)].astype(np.float64)+    labels = np.zeros(n, dtype=int)+    for _ in range(iters):+        d = ((xy[:, None, :] - cent[None, :, :]) ** 2).sum(-1) if n * k <= 4_000_000 else _chunked_d2(xy, cent)+        new = np.argmin(d, axis=1)+        if np.array_equal(new, labels):+            labels = new+            break+        labels = new+        for j in range(k):+            m = labels == j+            if m.any():+                cent[j] = xy[m].mean(0)+    return labels+++def _chunked_d2(xy: np.ndarray, cent: np.ndarray, chunk: int = 4096) -> np.ndarray:+    out = np.empty((xy.shape[0], cent.shape[0]), dtype=np.float64)+    for s in range(0, xy.shape[0], chunk):+        blk = xy[s:s + chunk]+        out[s:s + chunk] = ((blk[:, None, :] - cent[None, :, :]) ** 2).sum(-1)+    return out+++def spatial_mix_indices(labels_a, labels_b, coords_a, coords_b, t: float, n: int,+                        rng: np.random.Generator, cells_per_bin: int = 200, ndim: int = 2):+    """Like sample.mix_indices, but per shared type draw spatially stratified.++    coords_a / coords_b are the aligned (already rescaled) coordinates used+    for the output cloud; only their xy columns drive the binning.+    Returns (ia, ib, bin_labels) where bin_labels is an int array aligned to+    the concatenation of picked a-rows then b-rows (for diagnostics), or None.+    """+    n_b = int(np.clip(int(round(float(t) * n)), 0, n))+    n_a = n - n_b+    labels_a = np.asarray(labels_a).astype(str)+    labels_b = np.asarray(labels_b).astype(str)+    alloc_a = _alloc_by_type(labels_a, n_a) if n_a > 0 else {}+    alloc_b = _alloc_by_type(labels_b, n_b) if n_b > 0 else {}+    idx_a_by = {lab: np.flatnonzero(labels_a == lab) for lab in alloc_a}+    idx_b_by = {lab: np.flatnonzero(labels_b == lab) for lab in alloc_b}++    picks_a, picks_b = [], []+    shared = sorted(set(alloc_a) & set(alloc_b))+    for lab in sorted(set(alloc_a) | set(alloc_b)):+        ia_t = idx_a_by.get(lab, np.array([], dtype=int))+        ib_t = idx_b_by.get(lab, np.array([], dtype=int))+        k_a = int(min(alloc_a.get(lab, 0), ia_t.size))+        k_b = int(min(alloc_b.get(lab, 0), ib_t.size))+        pooled = ia_t.size + ib_t.size+        stratify = (lab in shared and pooled >= MIN_TYPE_CELLS+                    and k_a + k_b < pooled and cells_per_bin > 0)+        if not stratify:+            if k_a > 0:+                picks_a.append(ia_t if k_a == ia_t.size else rng.choice(ia_t, k_a, replace=False))+            if k_b > 0:+                picks_b.append(ib_t if k_b == ib_t.size else rng.choice(ib_t, k_b, replace=False))+            continue+        xy = np.vstack([coords_a[ia_t, :ndim], coords_b[ib_t, :ndim]]).astype(np.float64)+        k = max(4, pooled // max(int(cells_per_bin), 1))+        bins = _kmeans(xy, k, rng)+        ba = bins[:ia_t.size]+        bb = bins[ia_t.size:]+        wa = np.bincount(ba, minlength=k).astype(np.float64)+        wb = np.bincount(bb, minlength=k).astype(np.float64)+        sa = _spread(k_a, wa, wa.astype(int))+        sb = _spread(k_b, wb, wb.astype(int))+        for j in range(k):+            if sa[j] > 0:+                pool = ia_t[ba == j]+                picks_a.append(pool if sa[j] == pool.size else rng.choice(pool, int(sa[j]), replace=False))+            if sb[j] > 0:+                pool = ib_t[bb == j]+                picks_b.append(pool if sb[j] == pool.size else rng.choice(pool, int(sb[j]), replace=False))+    ia = np.concatenate(picks_a) if picks_a else np.array([], dtype=int)+    ib = np.concatenate(picks_b) if picks_b else np.array([], dtype=int)+    return ia, ib+++def interpolate_spatial(stage_a, stage_b, t: float, params: dict, cells_per_bin: int = 200, ndim: int = 2):+    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 = spatial_mix_indices(stage_a.labels, stage_b.labels, ca, cb, t, n, rng, cells_per_bin, ndim)++    expr_parts, coord_parts = [], []+    if ia.size:+        expr_parts.append(as_dense(stage_a.X, ia))+        coord_parts.append(np.asarray(ca[ia], dtype=np.float64))+    if ib.size:+        expr_parts.append(as_dense(stage_b.X, ib))+        coord_parts.append(np.asarray(cb[ib], dtype=np.float64))++    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),+        coord="mix", expr="carry", blend="convex", scale_damp=damp,+        nearer="b" if t >= 0.5 else "a",+        n_from_a=int(ia.size), n_from_b=int(ib.size),+        spatial_stratify=True, cells_per_bin=cells_per_bin,+    )+    return expr, coords.astype(np.float32), info

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

All verified. Final state: `solution/run.py` (mechanism on by default, bin=100) + `solution/spatial_mix.py` + `METHOD.md`.

**总结**

- **做了什么**：按 PLAN（T2HI-04）在父节点 mix 的型内随机抽样上实现空间分层抽样：共有型 xy k-means 分 bin（bin∈{100,200,400}、2D/3D 都试了），bin 内按 (1−t,t) 配比抽两阶段细胞；`SPATIAL_STRATIFY=0` 关闭对照与父节点同路径（逐位一致已验证）。机制确实生效：源阶段标签 Moran's I 0.302→0.287，kNN 同源边 0.665→0.659（理论 0.52）。
- **查分**：bin100 最优但 3 种子均值 58.99 vs 父 rank3 59.36，local_spatial 未提升（53.4–53.6 vs 54.0），shape_scale 略降。方向判负：proxy 上共有型只覆盖 20% 细胞，且真实 E8.75 的源阶段身份本就与空间位置相关，强行局部均匀化反而偏离。
- **下一步最值得试**：放弃抽样均匀化；local_spatial 瓶颈更可能在非共有型的空间占位（occupancy_dice/variogram），可试对非共有型按空间密度重加权抽样，或类型级坐标形状插值（改坐标生成而非选细胞）。

调研员

```json
{
  "name": "mix 空间分层抽样保邻域（spatial-stratified patch sampling）",
  "motivation": "父节点 2 四组分中 local_spatial 54.03、shape_scale 53.37 最弱。当前 mix 按型随机抽样混合两阶段细胞，同一型内空间上随机散布，导致输出云中局部邻域由不同阶段、不同空间位置的细胞拼凑，破坏邻域一致性。节点 5（T2HI-02 质心位移）已证明刚体位移对 local_spatial 无效（-0.58，噪声内）；节点 3（NN 匹配+坐标位移）生成失败。改变抽样策略（而非后处理位移）是尚未尝试的结构方向。",
  "approach": "在 mix 流程的 stratified sampling 步骤中，将每型内的随机抽样替换为空间分层抽样：\n1. 对齐+缩放后，对每个共有细胞型，用该型所有细胞（两阶段合并）的 xy 坐标做 k-means（k = max(4, n_type // 200)），得到空间 bin。\n2. 在每个 bin 内，按 (1-t, t) 比例分别从 stage_a、stage_b 抽样（四舍五入），保证每个空间区域都有正确比例的两阶段细胞。若某 bin 内某阶段细胞不足，从最近邻 bin 补。\n3. 表达和坐标随细胞一起走，与父节点相同。\n4. 非共有型（仅出现在一个阶段的型）保持原逻辑随机抽样。\n5. 单输入阶段退路（b is None）：无 bracket，直接走父节点的 fallback 路径，空间分层不适用，行为不变。\n6. 关键参数：k_per_type 初始 200 细胞/bin；搜索范围 [100, 400]。用 vec-score 跑 1 次确认方向，若 local_spatial 提升 >1 分再跑 2 次确认。\n7. 先小样本（取前 3000 细胞）验证抽样逻辑正确、无空 bin，再全量。",
  "expected_groups": ["local_spatial", "shape_scale"],
  "risks": "1) 空间 bin 太小导致某些 bin 内细胞数不足，抽样退化；Engineer 应打印每 bin 实际抽样数，若 >10% bin 不足则增大 bin 尺寸。2) 分层后表达分布可能微变，影响 expression_change；但每型总数不变，预期影响 <0.5 分。3) 若 local_spatial 的评分主要取决于全局形状而非邻域，则改善有限；用第 1 次查分即可判断。4) k-means 在极小型（<20 细胞）上不稳定，这些型退回随机抽样。",
  "family_id": "T2HI-04",
  "mechanism": "将 mix 的型内随机抽样改为空间分层抽样：每型按坐标 k-means 分 bin，每 bin 内按 (1-t, t) 比例从两阶段取样，保证输出云在每个空间区域都保留正确的两阶段混合比，从而维持局部邻域的空间一致性。",
  "vs_constant_shift": "不施加任何位移或全局变换；不改变表达值。改变的是哪些细胞被选入输出云——从型内随机改为空间分层，使局部邻域由空间相近的同阶段细胞组成，而非随机拼凑。这是抽样策略的结构变化，不是对已有云的坐标平移或缩放。",
  "mechanism_evidence": "1) 计算输出云中每个细胞的源阶段标签（a/b），再算该标签在输出坐标上的 Moran's I（空间自相关）；随机抽样下 I≈0，空间分层下每 bin 混合比一致，I 应更接近 0 或略负（无大团簇）。2) 计算输出云的 kNN 图中同源边比例，与理论值 t²+(1-t)² 比较。3) 四组分：预期 local_spatial 提升 >1 分，其余组分变化 <1 分。",
  "mechanism_off_control": "在 run.py 中设 SPATIAL_STRATIFY=False，退回父节点的型内随机抽样（同 seed），输出应与父节点逐位一致。对比两次查分：若 local_spatial 差异 <0.5 分，说明分层抽样未生效。"
}
```

审查员

REVIEW.json written