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

节点 n53

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 4399911e70: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 +7 −0、solution/README.md +4 −0、solution/run.py +62 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..006eac2--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,7 @@+## 改了什么+1. 修复 proxy2 崩溃根因（保留）：位移只用官方阶段（source != external）。原代码在 proxy2 上把外部 Qiu E9.0 当最新阶段抽样并计算跨数据集差值，导致输出变成心脏细胞+错误位移（proxy2 仅 27.4）。现在 proxy2 只有一个官方输入，退化为 copy_last。+2. 删除第 0/1 轮引入的 kNN 图扩散平滑：实测表明即使 blend=0.2、2 步，扩散仍严重塌缩局部结构（cell_state 32→7，covariation 22→2），对分数全面有害。回退到纯 copy_last / pseudobulk_shift。+## 用到的知识与出处+- 任务书 CONTRACT「外部输入阶段」一节：proxy2 的 Qiu E9.0 标签体系与技术不同，需按 source 区分。+- 方法卡「平移幅度」：常数位移在真实 T1 上低于 copy_last，proxy 单阶段时不强行位移。+- 第 1 轮实测：图扩散平滑全面破坏分数，已移除。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..66dfc8a--- /dev/null+++ b/solution/run.py@@ -0,0 +1,62 @@+#!/usr/bin/env python3+"""Graph-smoothed pseudobulk shift over official stages.++Delta is computed only from official stages (external inputs such as the+proxy2 Qiu E9.0 heart cells are excluded: their cell-type labels and+technology differ, so a cross-dataset delta destroys the prediction).+With fewer than two official stages the shift degenerates to copy_last.++Per the T1-09 direction, a light kNN graph diffusion (symmetric-normalised+heat kernel, few steps) is applied to the sampled output so neighbouring+cells become locally consistent while most per-cell variance is retained+(blend 0.2 targets the covariation score without collapsing cell_state).+"""++from __future__ import annotations++import argparse++import numpy as np+from src.task1_temporal.baselines import shift_rows, type_deltas+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,+)++++++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)+    official = [s for s in stages if s.get("source") != "external"]+    if not official:+        official = stages+    last = read_stage(args.data, official[-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]+    if len(official) >= 2:+        prev = read_stage(args.data, official[-2], genes)+        deltas = type_deltas(prev.X, labels_of(prev), last.X, labels_of(last))+        del prev+        X = shift_rows(X, labels_of(last)[rows], deltas)+    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

The graph diffusion is catastrophically destroying all scores (cell_state 32→7, covariation 22→2). Even at blend=0.2 with 2 steps, it collapses local structure. Remove it entirely; keep only the proxy2 fix.

<<<<<<< SEARCH
def graph_smooth(X, rng, k=15, steps=2, blend=0.2):
    from sklearn.decomposition import PCA
    from sklearn.neighbors import NearestNeighbors

    A = X.toarray() if sp.issparse(X) else np.asarray(X)
    A = np.asarray(A, dtype=np.float32)
    n = A.shape[0]
    if n < max(10, k + 1):
        return A
    Z = PCA(n_components=30, random_state=int(rng.integers(2**31))).fit_transform(A)
    nn = NearestNeighbors(n_neighbors=k, metric="cosine").fit(Z)
    _, idx = nn.kneighbors(Z)
    cols = idx[:, 1:].ravel()
    rows = np.repeat(np.arange(n), k - 1)
    G = sp.csr_matrix((np.ones(rows.size, dtype=np.float32),
                       (rows, cols)), shape=(n, n))
    G = (G + G.T).sign()
    deg = np.asarray(G.sum(axis=1)).ravel()
    deg[deg == 0] = 1.0
    D = sp.diags(1.0 / np.sqrt(deg))
    S = D @ G @ D
    S.setdiag(1.0)
    S = S.tocsr()
    F = A.copy()
    for _ in range(steps):
        F = S.dot(F)
    return np.asarray((1.0 - blend) * A + blend * F, dtype=np.float32)
=======
>>>>>>> REPLACE

<<<<<<< SEARCH
import scipy.sparse as sp

from src.task1_temporal.baselines import shift_rows, type_deltas
=======
from src.task1_temporal.baselines import shift_rows, type_deltas
>>>>>>> REPLACE