总览 · ← 返回运行 20261002-202908-search-t1-scr-D

节点 n4 在终选来历上

按细胞类型算 E(k-1)→E(k) 的基因均值差，做基因级 EB 收缩后按时间比例 α·r 只在非零元上加位移（保稀疏），外推到目标阶段。

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 1c84075fcf: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 915a381923。改动的文件：solution/EXECUTION.json +1 −0、solution/METHOD.md +53 −0、solution/README.md +4 −3、solution/run.py +119 −3

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..b80f125--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,53 @@+按细胞类型算 E(k-1)→E(k) 的基因均值差，做基因级 EB 收缩后按时间比例 α·r 只在非零元上加位移（保稀疏），外推到目标阶段。++## 方法（family: other / per-type damped displacement + EB shrinkage）++在 copy_last（父节点 1）基础上加机制：++1. 读最后两个输入阶段（视图无关：只遍历 `manifest["inputs"]`，按 `time` 排序；只有 1 个输入时严格退化为 copy_last）。+2. 对两阶段都出现的每个 `celltype` c，算逐基因伪批量差 δ_c,g = mean_g(stage2|c) − mean_g(stage1|c)（log 空间）。+3. 基因级经验贝叶斯收缩：λ = δ²/(δ² + σ²·(1/n1+1/n2))，σ² 为两阶段组内方差均值；收缩后 Δ = λ·δ。+4. stage2 独有类型用静态词表映射找 stage1 亲本（V-CM ← LV-CM+RV-CM 合并；Endocardium/BEC ← Endothelium），映射不到用全局（所有细胞）δ；这是官方 T1 词汇的改名/拆分知识（方法卡 §标签），不涉及任何保留阶段测量。+5. 输出细胞 = 最后阶段细胞抽样（同 copy_last，`target_n_cells`），每个细胞在其类型位移上做 x_new = max(0, x + α·r·Δ)，**只作用于已有非零元**（CSR data 原位修改），稀疏零结构不变。+6. r = (t_target − t_last)/(t_last − t_prev)，clip 到 [0,3]，只用相对时间差（视图平移不变）。X3 上 r=2（E9.0→E9.5 是 E8.75→E9.0 间隔的 2 倍），final 视图上 r=1，单输入视图 r 不适用（退化）。+7. α 默认 1.5（A 半上网格 {0,0.05,…,3} 选出，见下）。++## 关键实现发现++稠密加位移（对所有基因、含零元）会把 nnz 从 6% 涨到 28.5%，variogram 从 0.0016 涨到 0.0102，covariation 从 48.6 崩到 11.7（α=1 稠密版总分 42.6，**低于**对照）。改成只在非零元上加位移后 covariation 基本保住（45.3–47.6），总分大幅上升。这是本节点相对 PLAN 的主要修正。++## 查分记录（X3 A 半，seed 0，除非注明）++| 变体 | 总分 | de_rec | dir | cell_state | covar |+|---|---|---|---|---|---|+| 对照 α=0（=copy_last） | 47.62 | 43.09 | 49.12 | 49.51 | 48.59 |+| 稠密 α=0.5 | 43.54 | 49.53 | 49.37 | 50.94 | 17.67 |+| 稀疏 α=0.25 | 49.28 | 47.32 | 49.34 | 51.75 | 47.97 |+| 稀疏 α=0.5 | 50.79 | 50.48 | 49.44 | 54.28 | 47.61 |+| 稀疏 α=1.0 | 52.14 | 50.48 | 49.64 | 59.25 | 46.69 |+| **稀疏 α=1.5（提交）** | **52.91** | 50.48 | 49.62 | 62.76 | 45.32 |+| 稀疏 α=2.0 | 52.97 | 50.48 | 49.61 | 64.14 | 43.50 |+| 稀疏 α=3.0 | 51.16 | 50.48 | 49.62 | 60.95 | 39.23 |+| 稀疏 α=1.5, seed 1 | 53.80 | 50.00 | 51.20 | 64.64 | 45.55 |++α=1.5 与 2.0 差 0.06（噪声内），取 1.5 因 covariation 损失更小、更稳。相对对照 +5.3，远超 2 分噪声。++## 机制生效证据（mechanism_off_control）++- 对照（α=0，同一程序）：47.62 ≈ copy_last 47.92（B 半 vs A 半差异，在 0.5 容差内），确认管线本身无偏。+- λ 分布有区分度：各类型 f(λ>0.5) 在 0.08–0.24，中位 λ≈0（多数基因被收缩掉，非全 1/全 0）。+- 每类型 |λδ|>0.1 的基因数 368–3916（主类型 IFT-CM 737、Endocardium 584、Unknown 368），位移非平凡。+- 四组分变化（α=0 → 1.5）：de_recovery 43.1→50.5，cell_state 49.5→62.8，direction 49.1→49.6，covariation 48.6→45.3（−3.3，超过 PLAN 预期的 −1，但总分净增；α=1.0 时 covariation 只 −1.9）。+- 输出稀疏度不变：nnz 比例 6.05%（对照 6.09%），细胞确实被移动（mmd_u 0.0344→0.0241，向真值靠近）。++## 验证过 / 没验证++- 验证：X3 视图 vec-check ok；seed 0 双跑逐元素一致（确定）；seed 1 分数 53.80；单输入合成视图（只留 input_0）跑通且 vec-check ok（copy_last 退路）；`EXECUTION.json {"gpu": false}`（纯 CPU，秒级）。+- 没验证：final 视图（E8.5+E9.5→E10.5，r=1，有效位移只有 X3 的一半，α 最优值可能不同——但节点分数只按 X3）；proxy2；B 半真值。+- α 是在 X3 A 半上选的，A/B 半差异可能让最优点漂移，但 α∈[1.0,2.0] 都在 52+ 的平台区，稳健。++## 知识来源++- 标签改名/拆分（Endothelium→Endocardium/BEC，LV/RV-CM→V-CM）：任务书方法卡「T1 卡：E10.5 v1 §标签」，属通用词汇知识，不含保留阶段测量。+- EB 收缩公式 δ²/(δ²+σ²/n)：标准经验贝叶斯（PLAN 指定）。+- 未使用 external/、prior/ 数据；未使用任何禁窗内信息。diff --git a/solution/README.md b/solution/README.mdindex e6cdf6f..46d99f4 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,4 +1,5 @@-# copy_last+# per-type EB-shrunk displacement (sparse-only) -输出最新一个输入阶段（proxy：E8.5；final：E9.5），随机无放回抽到榜的细胞上限（T1 5118）。-不改表达，不改组成。这是地板：proxy 预期 ≈ 50（seed 0 实测 49.77；抽到 5118 个细胞，比整份 E8.5 的 50.00 略低）。+copy_last 基础上：按 celltype 算最后两个输入阶段的基因均值差，EB 收缩（λ=δ²/(δ²+σ²(1/n1+1/n2))），+乘 α·r（α 默认 1.5，r=外推时间/输入间隔）后只加在输出细胞的已有非零元上（保稀疏结构），max(0,·) 截断。+单输入视图严格退化为 copy_last；`--alpha 0` 为机制关闭对照。详见 METHOD.md。diff --git a/solution/run.py b/solution/run.pyindex aa4bbf0..8903ba9 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,14 +1,33 @@ #!/usr/bin/env python3-"""copy_last: the latest input stage, subsampled into the board's cell range."""+"""Per-type damped displacement with gene-level empirical-Bayes shrinkage.++Base: copy_last (output = cells of the latest input stage, subsampled).+Mechanism: for every cell type present in the two latest input stages, compute+the per-gene pseudobulk mean shift delta_g = mean(stage2|c) - mean(stage1|c) in+log space, shrink it with gene-level EB weights+    lambda_g = delta_g^2 / (delta_g^2 + sigma_g^2 (1/n1 + 1/n2)),+and move each output cell by  x_new = max(0, x + alpha * r * lambda * delta),+where r = (t_target - t_last) / (t_last - t_prev) rescales the observed step to+the extrapolation horizon (relative times only -> view independent).++Types present only in stage2 are matched to stage1 parents via a static+label-vocabulary map (general lineage knowledge, see METHOD.md); unmatched+types fall back to the global (all-cell) pseudobulk delta.+With a single input stage the program reduces exactly to copy_last.+--alpha 0 disables the mechanism entirely (control, == copy_last).+"""  from __future__ import annotations  import argparse+import os  import numpy as np+from scipy import sparse  from src.task1_temporal.view_io import (     inputs_by_time,+    labels_of,     load_manifest,     panel_genes,     read_stage,@@ -17,20 +36,117 @@ from src.task1_temporal.view_io import (     write_prediction, ) +# Static vocabulary knowledge (label renames/splits in the official T1 lexicon,+# general lineage knowledge, not derived from held-out measurements):+# Endothelium splits into Endocardium / BEC; LV-CM + RV-CM are renamed V-CM.+SYNONYM_PARENTS = {+    "V-CM": ("LV-CM", "RV-CM"),+    "Endocardium": ("Endothelium",),+    "BEC": ("Endothelium",),+}++CAP_R = 3.0+++def group_means(X: sparse.csr_matrix, labels: np.ndarray, groups: dict[str, np.ndarray]):+    """Per-group mean and mean-of-squares over genes (sparse-friendly)."""+    out = {}+    X2 = X.multiply(X).tocsr()+    for name, mask in groups.items():+        n = int(mask.sum())+        if n == 0:+            continue+        idx = np.flatnonzero(mask)+        m1 = np.asarray(X[idx].mean(axis=0), dtype=np.float64).ravel()+        m2 = np.asarray(X2[idx].mean(axis=0), dtype=np.float64).ravel()+        var = np.maximum(m2 - m1 * m1, 0.0)+        out[name] = (n, m1, var)+    return out+  def main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)     parser.add_argument("--out", required=True)     parser.add_argument("--seed", type=int, default=0)+    parser.add_argument("--alpha", type=float,+                        default=float(os.environ.get("VEC_ALPHA", "1.5")))+    parser.add_argument("--no-time-scale", action="store_true",+                        default=os.environ.get("VEC_NO_TIME_SCALE", "") == "1")     args = parser.parse_args()      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)-    last = read_stage(args.data, inputs_by_time(manifest)[-1], genes)+    entries = inputs_by_time(manifest)     rng = np.random.default_rng(args.seed)++    last = read_stage(args.data, entries[-1], genes)     rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)-    write_prediction(last.X[rows], genes, args.out, seed=args.seed)++    if len(entries) >= 2 and args.alpha != 0.0:+        prev = read_stage(args.data, entries[-2], genes)+        lab1 = labels_of(prev)+        lab2 = labels_of(last)+        types1 = np.unique(lab1)+        types2 = np.unique(lab2)++        groups1 = {"__global__": np.ones(prev.n_obs, dtype=bool)}+        for t in types1:+            groups1[str(t)] = lab1 == t+        groups2 = {"__global__": np.ones(last.n_obs, dtype=bool)}+        for t in types2:+            groups2[str(t)] = lab2 == t++        s1 = group_means(prev.X.tocsr(), lab1, groups1)+        s2 = group_means(last.X.tocsr(), lab2, groups2)++        def delta_for(name2: str):+            """(shrunk delta vector) for a stage-2 group."""+            n2, m2, v2 = s2[name2]+            if name2 in s1:+                n1, m1, v1 = s1[name2]+            else:+                parents = [p for p in SYNONYM_PARENTS.get(name2, ()) if p in s1]+                if parents:+                    mask = np.zeros(prev.n_obs, dtype=bool)+                    for p in parents:+                        mask |= lab1 == p+                    pooled = group_means(prev.X.tocsr(), lab1, {"_p": mask})["_p"]+                    n1, m1, v1 = pooled+                else:+                    n1, m1, v1 = s1["__global__"]+            d = m2 - m1+            noise = 0.5 * (v1 + v2) * (1.0 / n1 + 1.0 / n2)+            lam = np.where(d * d + noise > 0, d * d / np.maximum(d * d + noise, 1e-12), 0.0)+            return lam * d++        # time rescale from relative times only+        r = 1.0+        if not args.no_time_scale:+            t_prev = float(entries[-2]["time"])+            t_last = float(entries[-1]["time"])+            t_tgt = float(manifest["target"]["time"])+            if t_last > t_prev:+                r = float(np.clip((t_tgt - t_last) / (t_last - t_prev), 0.0, CAP_R))+        scale = args.alpha * r++        # shift only the sampled output cells, at existing nonzeros only+        # (preserves the sparse zero structure -> covariation intact)+        Xs = last.X[rows].tocsr()+        labs = lab2[rows]+        for t in np.unique(labs):+            d = (scale * delta_for(str(t))).astype(np.float32)+            tmask = np.flatnonzero(labs == t)+            for i in tmask:+                lo, hi = Xs.indptr[i], Xs.indptr[i + 1]+                if hi > lo:+                    cols = Xs.indices[lo:hi]+                    Xs.data[lo:hi] = np.maximum(Xs.data[lo:hi] + d[cols], 0.0)+        X = Xs+    else:+        X = last.X[rows]++    write_prediction(X, genes, args.out, seed=args.seed)   if __name__ == "__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

**完成。** 节点 4（improve，父 copy_last 47.92）：

- **做了什么**：实现 PLAN 的 per-type 阻尼位移 + 基因级 EB 收缩；关键修正是**位移只作用于已有非零元**（稠密版把 covariation 从 48.6 砸到 11.7、总分低于对照），并按 r=(目标−末输入)/(输入间隔) 做时间缩放。α 网格选出 1.5；stage2 独有类型用静态改名词表映射亲本，单输入视图退化为 copy_last，`--alpha 0` 为机制关闭对照。
- **查分（X3 A 半，14/20 次）**：对照 α=0 → 47.62；提交版 α=1.5 → **52.91**（seed 1 复验 53.80），de_recovery 43→50、cell_state 49→63、covariation 仅 −3.3。vec-check ok、双跑逐元素确定、耗时 ~1s、纯 CPU（EXECUTION.json gpu:false）。
- **下一步最值得试**：① covariation 是唯一净损项——可试对位移做低秩/通路级平滑（prior/ Reactome）减少逐基因噪声；② final 视图 r=1 时有效位移减半，α 或需按 r 分段校准；③ 与节点 2 的 OT 耦合（50.65）组合：OT 补组成变化，EB 位移补表达外推。

调研员

 to global (type-agnostic) per-gene mean shift. 2) Optimal α may be ≈0, meaning shrinkage kills all signal — detect early: if α=0.2 shows no improvement over copy_last on first vec-score query, abort and report mechanism failure. 3) X3 is heart-only data with potentially fewer types; small n_c inflates σ², over-shrinking real signal — mitigate by pooling types with n_c < 20 into a 'rest' group. 4) Score noise ~2 pts: require two consistent vec-score queries before concluding improvement.",
  "family_id": "other",
  "mechanism": "Per-cell-type mean displacement between consecutive stages, projected forward one step with gene-level empirical Bayes shrinkage that attenuates noisy gene deltas proportional to their estimation uncertainty (δ²/(δ²+σ²/n)).",
  "vs_constant_shift": "Old pseudobulk_shift applies a uniform alpha=1 per-type delta to all genes equally. This variant adds (a) gene-level shrinkage λ_c,g so unreliable deltas are suppressed toward zero while consistent DE genes retain their shift, and (b) a global alpha < 1 to prevent overshooting. The shrinkage is data-driven (EB), not a hand-tuned constant, and the mechanism specifically targets the signal-to-noise ratio of each gene's temporal change within each type.",
  "mechanism_evidence": "Engineer should report: (1) distribution of λ values — if all λ≈1 or all λ≈0, shrinkage is not differentiating; (2) number of genes with |shrunk_delta| > 0.1 per type (should be non-trivial, e.g. 50–500); (3) four-component scores at α=0 vs best α — de_recovery and direction should rise while cell_state and covariation should not drop >1 pt; (4) per-type mean shift magnitude in PCA space to confirm cells actually moved.",
  "mechanism_off_control": "Run the same script with --alpha 0.0 (or --no-shift flag). This skips step 5 entirely, producing output identical to copy_last. Expected: score ≈ 47.92, de_recovery ≈ 44.09. Any deviation > 0.5 pt indicates a bug in the pipeline rather than the mechanism.",
  "sources": []
}
```

审查员

REVIEW.json written