总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-C-native

节点 n16

改了什么

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 deef55e6f3: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/METHOD.md +11 −0、solution/README.md +4 −0、solution/run.py +105 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..6b148b8--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,11 @@+## 改了什么+1. 用 `inputs_by_time(manifest, include_external=False)` 只取官方阶段作为采样基础，修复 proxy2 中从 Qiu E9.0（仅心脏细胞、标签不匹配）采样导致 cell_state/covariation 崩溃的问题。+2. 实现按类型伪批量位移 + 全局收缩系数 α=0.3（官方 α=1 得分 48.6 低于 copy_last ~50，故取保守值）。+3. 增加逐基因 James-Stein 经验贝叶斯收缩：对每个基因计算 delta 的标准误（基于两组细胞内方差/样本量），收缩因子 = delta²/(delta²+SE²)，噪声大的基因位移更小。+4. 对数空间操作，结果夹到 ≥0。无同名前一阶段的类型保持原样。+5. 单输入时退化为 copy_last（与 proxy/proxy2 兼容）。+## 用到的知识与出处+- k018：收缩系数 α 的必要性，官方 α=1 低于地板；逐基因经验贝叶斯收缩建议。+- k012：T1 评分组（DE 25%、direction 25%、cell_state 30%、covariation 20%），copy_last=50 为地板。+- 方向库 T1-02 描述：对数空间位移、夹到 ≥0、无同名类型保持原样。+- CONTRACT proxy2 说明：外部输入阶段细胞类型名与官方不同，不能直接配对。diff --git a/solution/README.md b/solution/README.mdnew file mode 100644index 0000000..ba29577--- /dev/null+++ b/solution/README.md@@ -0,0 +1,4 @@+# pseudobulk_shift++最新阶段抽样后，每个细胞加上所属类型在最后一步的伪批量差值 mean(last|type) − mean(prev|type)，夹到 ≥0；前一阶段没有的类型原样复制。+T1 proxy 只有一个输入阶段，没有差值可取，退化成 copy_last（同样的抽样），所以 proxy 分 = copy_last（seed 0 实测 49.77）。final 才真正平移；官方在真实 T1 上报的常数位移是 48.6，低于地板。diff --git a/solution/run.py b/solution/run.pynew file mode 100644index 0000000..289e4fb--- /dev/null+++ b/solution/run.py@@ -0,0 +1,105 @@+#!/usr/bin/env python3+"""Per-type pseudobulk shift with shrinkage alpha and per-gene empirical Bayes.++Samples from the last official (non-external) stage. When 2+ official stages+exist, computes per-type delta = mean(last|c) - mean(prev|c) in log space,+applies global alpha and per-gene James-Stein shrinkage, then adds to each+cell (clipped >= 0). Types absent from the earlier stage are copied unchanged.++With a single official input (proxy, proxy2), falls back to copy_last.+External inputs are excluded from base sampling to avoid cell-type mismatch.+"""++from __future__ import annotations++import argparse++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,+    sample_rows,+    target_n_cells,+    write_prediction,+)++ALPHA = 0.3+++def _to_dense(X):+    if sparse.issparse(X):+        return np.asarray(X.toarray(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++def compute_shrunk_deltas(X_last, labels_last, X_prev, labels_prev, alpha):+    Xl = _to_dense(X_last)+    Xp = _to_dense(X_prev)+    cats_prev = set(np.unique(labels_prev))+    deltas = {}+    for c in np.unique(labels_last):+        if c not in cats_prev:+            continue+        ml = labels_last == c+        mp = labels_prev == c+        nl, np_ = ml.sum(), mp.sum()+        if nl < 2 or np_ < 2:+            continue+        mean_l = Xl[ml].mean(axis=0)+        mean_p = Xp[mp].mean(axis=0)+        delta = mean_l - mean_p+        var_l = Xl[ml].var(axis=0)+        var_p = Xp[mp].var(axis=0)+        se2 = var_l / nl + var_p / np_+        shrink = delta**2 / (delta**2 + se2 + 1e-10)+        deltas[c] = (alpha * shrink * delta).astype(np.float32)+    return deltas+++def apply_deltas(X, labels, deltas):+    Xd = _to_dense(X)+    for c, d in deltas.items():+        mask = labels == c+        if mask.any():+            Xd[mask] += d+    np.clip(Xd, 0, None, out=Xd)+    return Xd+++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)+    args = parser.parse_args()++    manifest = load_manifest(args.data)+    genes = panel_genes(args.data, manifest)+    stages = inputs_by_time(manifest, include_external=False)+    last = read_stage(args.data, stages[-1], genes)+    rng = np.random.default_rng(args.seed)+    rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)+    X = last.X[rows]+    labels = labels_of(last)++    if len(stages) >= 2:+        prev = read_stage(args.data, stages[-2], genes)+        deltas = compute_shrunk_deltas(+            last.X, labels, prev.X, labels_of(prev), ALPHA+        )+        del prev+        if deltas:+            X = apply_deltas(X, labels[rows], deltas)+    else:+        X = _to_dense(X)++    write_prediction(X, genes, args.out, seed=args.seed)+++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）。

没有记录调研来源。

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

native

stream

<<<<<<< SEARCH
ALPHA = 0.35
PROLIF_GENES = ["Mki67", "Top2a", "Pcna"]
PROLIF_BETA = 0.15
=======
ALPHA = 0.3
>>>>>>> REPLACE

<<<<<<< SEARCH
def proliferation_weights(adata, genes, labels, beta):
    gene_list = list(genes)
    indices = [gene_list.index(g) for g in PROLIF_GENES if g in gene_list]
    if not indices:
        return None
    X = _to_dense(adata.X)
    prolif = X[:, indices].mean(axis=1)
    cats = np.unique(labels)
    type_scores = {}
    for c in cats:
        mask = labels == c
        if mask.sum() > 0:
            type_scores[c] = float(prolif[mask].mean())
    if len(type_scores) < 2:
        return None
    overall_mean = np.mean(list(type_scores.values()))
    weights = np.ones(adata.n_obs, dtype=np.float64)
    for c, score in type_scores.items():
        mask = labels == c
        rel = score - overall_mean
        weights[mask] = np.exp(beta * rel)
    weights /= weights.sum()
    return weights


def weighted_sample(n, target, weights, rng):
    if target >= n:
        return rng.choice(n, size=target, replace=True, p=weights)
    return rng.choice(n, size=target, replace=False, p=weights)
=======
def stratified_sample(labels, n_target, rng):