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

节点 n28

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 7ea199fddb: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 34254aa790。改动的文件：solution/METHOD.md +2 −2、solution/run.py +54 −2

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 58fac39..3479e6d 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,4 +1,4 @@ ## 改了什么-从 pseudobulk_shift 改为组成重加权（T1-01 方向）。核心变化：(1) 使用 include_external=False 避免 proxy2 中 Qiu E9.0 标签不匹配导致的崩溃（原程序 proxy2 仅 27.43）；(2) 按正则模式将细胞类型分为 heart（×1.6）、reduced（×0.15，含 surface ectoderm/EXEM/paraxial/neural）、gut（×0.9）、neutral（×1.0）四族，用加权抽样替代均匀抽样，不改表达值；(3) 对 X3 等无官方输入的视图有退路（回退到全部输入）。权重规则按名称模式定义，不写死特定阶段的名字。+相对父节点（节点9）：(1) 恢复父节点的逐细胞加权有放回抽样（第0轮改为分层抽样导致 covariation 从 51.96 暴跌到 47.30，证明分层抽样破坏了经验协方差结构）；(2) 当有两个官方输入阶段时（final 视图），叠加极保守的经验贝叶斯收缩伪批量位移（α=0.05，比第0轮的0.15更小），目标是温和提升 de_recovery 而不损害其他分组；单阶段视图（proxy/proxy2/X3）行为与父节点完全一致。第0轮教训：分层抽样虽然控制了类型比例，但破坏了细胞间基因共变结构，covariation 损失远超 de_recovery 收益。 ## 用到的知识与出处-方向库 T1-01（种子 heart_jcf_peri：心脏类 ×1.6，表面外胚层/EXEM/轴旁中胚层 ×0.25，丢 Neural Tube；proxy 55.97）；方法卡 T1 标签节（E8.5 独有名应降权或丢弃）；k012（cell_state 权重 30%，组成匹配是关键得分项）。+方向库 T1-01（组成重加权）、T1-02（按类型伪批量位移+收缩系数）；方法卡 k018（α∈[0,1]，官方 α=1 低于 copy_last）；第0轮实验反馈（分层抽样导致 covariation -4.66，α=0.15 位移仅带来 de_recovery +1.0，净效果为负）。diff --git a/solution/run.py b/solution/run.pyindex 4710196..2996919 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,9 +1,12 @@ #!/usr/bin/env python3-"""Composition reweighting: weighted sampling from latest official input stage.+"""Composition reweighting with per-cell weighted sampling + tiny EB delta.  Cell types are classified into families by regex patterns on their names. Heart-related types are upweighted, surface ectoderm / extraembryonic /-neural types are downweighted or dropped. Expression values are unchanged.+neural types are downweighted. Per-cell weighted sampling with replacement+preserves empirical covariation structure. When two official stages are+available, a very small empirical-Bayes shrunk pseudobulk delta (alpha=0.05)+is applied to same-name types to nudge de_recovery.  Uses include_external=False to avoid mismatched-label external stages in proxy2. Falls back to all stages if no official inputs exist (X3).@@ -15,6 +18,7 @@ import argparse import re  import numpy as np+from scipy import sparse as sp  from src.task1_temporal.view_io import (     inputs_by_time,@@ -40,6 +44,8 @@ _GUT_PAT = re.compile(     re.IGNORECASE, ) +ALPHA = 0.05+  def family_weight(name: str) -> float:     if _HEART_PAT.search(name):@@ -51,6 +57,36 @@ def family_weight(name: str) -> float:     return 1.0  +def _dense(X):+    if sp.issparse(X):+        return np.asarray(X.todense())+    return np.asarray(X)+++def eb_deltas(prev_X, prev_labels, last_X, last_labels, alpha):+    prev_types = set(np.unique(prev_labels))+    deltas = {}+    for t in np.unique(last_labels):+        if t not in prev_types:+            deltas[t] = None+            continue+        lm = last_labels == t+        pm = prev_labels == t+        nl, np_ = int(lm.sum()), int(pm.sum())+        if nl < 3 or np_ < 3:+            deltas[t] = None+            continue+        mean_l = _dense(last_X[lm]).mean(axis=0).ravel()+        mean_p = _dense(prev_X[pm]).mean(axis=0).ravel()+        var_l = _dense(last_X[lm]).var(axis=0).ravel()+        var_p = _dense(prev_X[pm]).var(axis=0).ravel()+        delta = mean_l - mean_p+        se2 = var_p / max(np_, 1) + var_l / max(nl, 1)+        shrink = delta**2 / (delta**2 + se2 + 1e-10)+        deltas[t] = alpha * shrink * delta+    return deltas++ def main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)@@ -77,6 +113,22 @@ def main() -> None:      rows = rng.choice(last.n_obs, size=n_target, replace=True, p=weights)     X = last.X[rows]++    if len(stages) >= 2:+        prev = read_stage(args.data, stages[-2], genes)+        deltas = eb_deltas(prev.X, labels_of(prev), last.X, labels, ALPHA)+        del prev+        X = _dense(X)+        sel_labels = labels[rows]+        for t, d in deltas.items():+            if d is None:+                continue+            mask = sel_labels == t+            if mask.any():+                X[mask] = np.clip(X[mask] + d, 0, None)+        from scipy.sparse import csr_matrix+        X = csr_matrix(X)+     write_prediction(X, genes, args.out, 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）。

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

Looking at the results: Round 1 score 50.98 vs parent 51.40. The EB delta is hurting X3 (45.85 vs 47.06) and covariation (49.07 vs 51.96) without helping de_recovery. The delta only fires on multi-stage views (X3, final) and is net negative.

Strategy: Remove delta entirely, tune weights more aggressively (heart up, reduced down), add NCC to heart family (cardiac NCC contributes to outflow tract at these stages).

<<<<<<< SEARCH
#!/usr/bin/env python3
"""Composition reweighting with per-cell weighted sampling + tiny EB delta.

Cell types are classified into families by regex patterns on their names.
Heart-related types are upweighted, surface ectoderm / extraembryonic /
neural types are downweighted. Per-cell weighted sampling with replacement
preserves empirical covariation structure. When two official stages are
available, a very small empirical-Bayes shrunk pseudobulk delta (alpha=0.05)
is applied to same-name types to nudge de_recovery.

Uses include_external=False to avoid mismatched-label external stages in
proxy2. Falls back to all stages if no official inputs exist (X3).
"""

from __future__ import annotations

import argparse
import re

import numpy as np
from scipy import sparse as sp

from src.task1_temporal.view_io import (
    inputs_by_time,
    labels_of,
    load_manifest,
    panel_genes,
    read_stage,
    target_n_cells,
    write_prediction,
)

_HEART_PAT = re.compile(
    r"cm\b|cardi|heart|shf|endocard|oft\b|avc|ift\b|sv[- ]cm|v[- ]cm|"
    r"proepicard|pericard|epicard|blood",