总览 · ← 返回运行 20261001-233757-search-t1-abc-r0-C-native

节点 n45

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 9b3a7b38e1: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 36251e2337。改动的文件：solution/METHOD.md +7 −9、solution/run.py +21 −58

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 26f9540..06ba225 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,12 +1,10 @@ ## 改了什么-相对父节点的全部改动：-1. 将全局加权抽样替换为按细胞类型分层抽样（stratified sampling）：先按 family_weight(type) × n_cells(type) 计算每个类型的配额分数，按比例分配整数配额（floor + 余量贪心补齐），再在每个类型内无放回均匀抽样。这消除了全局加权抽样中类型配额随机波动的问题，保证每个类型获得确定性的细胞数，从而更精确保留类内基因共变结构（covariation 是最弱维度 41.31）。-2. 类型配额分配使用 floor + 贪心补齐策略，确保总细胞数精确等于 n_target 且不超过各类型可用细胞数。-3. 非官方输入（X3）分支保持简单均匀抽样不变。-4. 两阶段伪批量位移逻辑不变。-5. 新增：将 n_target 夹到 manifest 的 [min_cells, max_cells] 范围内，修复 target_n_cells 返回值超过 max_cells 导致 invalid_format 的问题（5127 > 5118）。+1. 恢复 family_weight 加权重写（cardiac ×1.6, downweight ×0.25, Neural Tube ×0），因为第 1 轮移除后所有维度均大幅下降（proxy 53.52→49.93），证明加权对 cell_state 和 direction 有显著正贡献。+2. 将第 0 轮的分层抽样（stratified）改回全局加权抽样（每个细胞按其类型的 family_weight 获得概率权重，用 rng.choice 的 p 参数一次抽取）。理由：分层抽样强制每个类型精确配额，消除了类型比例的随机波动，而这种波动本身是群体协方差结构的一部分；全局加权抽样保留了这种自然变异，预期改善 covariation。+3. 保留 n_target clamp 修复和 assert 断言。+4. 非官方输入（X3）分支保持简单均匀抽样。 ## 用到的知识与出处-- 父节点 ANALYSIS 建议：按细胞类型分层抽样（每类型按权重分配整数配额再在类内无放回抽取），保留类内共变结构-- k012：covariation 占 20%，是最弱维度（41.31），改善空间最大-- 方法卡 T1-01 方向：族权重参数不变（心脏 ×1.6，表面外胚层/EXEM/轴旁 ×0.25，Neural Tube 丢弃）+- 实验表：节点 27（全局加权抽样+位移）得 proxy 54.49，是本家族最高；第 0 轮分层抽样降至 53.52，第 1 轮无加权降至 49.93+- k012：covariation 占 20%，全局加权抽样保留类型比例的自然波动，有利于协方差结构+- k036：组成预测中权重收缩；此处不收缩（α=1 等效），因为实验证明加权有效 - CONTRACT：输出细胞数必须在 [min_cells, max_cells] 之内diff --git a/solution/run.py b/solution/run.pyindex b3e5777..0370fda 100644--- a/solution/run.py+++ b/solution/run.py@@ -50,61 +50,18 @@ def family_weight(name: str) -> float:     return DEFAULT_W  -def stratified_sample(labels, n_target, rng, weight_fn=None):-    types = np.unique(labels)-    type_indices = {}-    type_scores = {}-    for t in types:-        idx = np.where(labels == t)[0]-        type_indices[t] = idx-        w = weight_fn(t) if weight_fn else 1.0-        type_scores[t] = w * len(idx)--    total_score = sum(type_scores.values())-    if total_score <= 0:-        n = min(n_target, len(labels))-        return rng.choice(len(labels), size=n, replace=False)--    sorted_types = sorted(types, key=lambda t: type_scores[t], reverse=True)-    quotas = {}-    remaining = n_target-    for i, t in enumerate(sorted_types):-        n_avail = len(type_indices[t])-        if i == len(sorted_types) - 1:-            quotas[t] = min(remaining, n_avail)-        else:-            q = int(np.floor(n_target * type_scores[t] / total_score))-            q = max(q, 0)-            q = min(q, n_avail, remaining)-            quotas[t] = q-            remaining -= q--    if remaining > 0:-        for t in sorted_types:-            if remaining <= 0:-                break-            n_avail = len(type_indices[t])-            can_add = n_avail - quotas.get(t, 0)-            add = min(can_add, remaining)-            if add > 0:-                quotas[t] = quotas.get(t, 0) + add-                remaining -= add--    row_list = []-    for t in types:-        q = quotas.get(t, 0)-        if q <= 0:-            continue-        idx = type_indices[t]-        if q >= len(idx):-            row_list.append(idx)-        else:-            row_list.append(rng.choice(idx, size=q, replace=False))--    if not row_list:-        n = min(n_target, len(labels))-        return rng.choice(len(labels), size=n, replace=False)-    return np.concatenate(row_list)+def weighted_sample(labels, n_target, rng):+    n_total = len(labels)+    weights = np.array([family_weight(t) for t in labels], dtype=np.float64)+    total_w = weights.sum()+    if total_w <= 0:+        if n_target <= n_total:+            return rng.choice(n_total, size=n_target, replace=False)+        return rng.choice(n_total, size=n_target, replace=True)+    probs = weights / total_w+    if n_target <= n_total:+        return rng.choice(n_total, size=n_target, replace=False, p=probs)+    return rng.choice(n_total, size=n_target, replace=True, p=probs)   def has_official_stage(manifest) -> bool:@@ -130,11 +87,16 @@ def main() -> None:      labels = labels_of(last)     n_target = target_n_cells(manifest, last.n_obs)-    n_target = min(n_target, manifest.get("max_cells", n_target))-    n_target = max(n_target, manifest.get("min_cells", n_target))+    max_c = manifest.get("max_cells", None)+    min_c = manifest.get("min_cells", None)+    if max_c is not None:+        n_target = min(n_target, max_c)+    if min_c is not None:+        n_target = max(n_target, min_c)+    n_target = max(n_target, 1)      if official:-        rows = stratified_sample(labels, n_target, rng, weight_fn=family_weight)+        rows = weighted_sample(labels, n_target, rng)     else:         if n_target <= last.n_obs:             rows = rng.choice(last.n_obs, size=n_target, replace=False)@@ -148,6 +110,7 @@ def main() -> None:         del prev         X = shift_rows(X, labels[rows], deltas) +    assert X.shape[0] == n_target, f"Expected {n_target} cells, got {X.shape[0]}"     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, removing family weights hurt everything dramatically (proxy dropped from 53.52 to 49.93). The weights were clearly beneficial. I need to restore them but use global weighted sampling (like node 27 which scored proxy 54.49) instead of stratified sampling, which should better preserve covariation.

<<<<<<< SEARCH
#!/usr/bin/env python3
"""Composition reweighting by cell-type family + pseudobulk shift.

Reweights cells from the latest official input stage by cardiac vs non-cardiac
family membership before sampling, then applies per-type pseudobulk delta if
two official stages are available. External inputs are excluded from sampling
base to avoid proxy2 collapse onto cardiac-only Qiu cells.
"""

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,
    target_n_cells,
    write_prediction,
)




def proportional_sample(labels, n_target, rng):
    n_total = len(labels)
    if n_target <= n_total:
        return rng.choice(n_total, size=n_target, replace=False)
    return rng.choice(n_total, size=n_target, replace=True)


def has_official_stage(manifest) -> bool: