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

节点 n10

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 9e4707bac4: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 5c87608489。改动的文件：solution/METHOD.md +11 −0、solution/run.py +28 −4

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..b2fcd70--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,11 @@+## 改了什么+相对父节点 copy_last 的全部改动：+1. 去掉了第 0、1 轮的加性位移逻辑（实验证明任何均值位移都严重损害 covariation：alpha=0.03 导致 48.44→40.79，alpha=0.008 仍导致 48.44→44.93，且 de_recovery 和 direction 均未改善）+2. 改用 max_cells 而非 target_n_cells 决定输出细胞数：更多细胞使协方差估计更稳定，可能改善 covariation+3. 改用按类型分层抽样（stratified sampling）替代随机抽样：精确保持类型比例，减少随机抽样引入的组成噪声，可能改善 cell_state+4. 去掉了 scipy.sparse 的 import（不再需要稠密化）+本质上仍是 copy_last（不改表达），但通过更优的抽样策略和更多细胞数来略微改善评分。+## 用到的知识与出处+- k012：covariation 占 20%，对基因共变结构敏感；更多细胞改善协方差估计+- k016：检查组成保留和稀有状态保留；分层抽样更好保留稀有类型+- 第 0、1 轮实验反馈：加性位移无论多小都损害 covariation，收益不抵成本diff --git a/solution/run.py b/solution/run.pyindex aa4bbf0..dd164c5 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,5 +1,5 @@ #!/usr/bin/env python3-"""copy_last: the latest input stage, subsampled into the board's cell range."""+"""Stratified copy_last: output max cells with exact type-proportional sampling."""  from __future__ import annotations @@ -17,7 +17,6 @@ from src.task1_temporal.view_io import (     write_prediction, ) - def main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)@@ -27,9 +26,34 @@ def main() -> None:      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)+    last = read_stage(args.data, entries[-1], genes)     rng = np.random.default_rng(args.seed)-    rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)++    n_max = manifest.get("max_cells", last.n_obs)+    n_out = min(last.n_obs, n_max)++    labels = last.obs["celltype"].values+    unique_types = np.unique(labels)+    type_counts = np.array([(labels == t).sum() for t in unique_types])+    props = type_counts / type_counts.sum()+    allocations = np.floor(props * n_out).astype(int)+    remainder = n_out - allocations.sum()+    if remainder > 0:+        frac = props * n_out - allocations+        top_idx = np.argsort(-frac)[: int(remainder)]+        allocations[top_idx] += 1++    rows_list = []+    for i, t in enumerate(unique_types):+        type_idx = np.where(labels == t)[0]+        n_sample = min(int(allocations[i]), len(type_idx))+        if n_sample > 0:+            chosen = rng.choice(type_idx, size=n_sample, replace=False)+            rows_list.append(chosen)+    rows = np.concatenate(rows_list)+    rng.shuffle(rows)+     write_prediction(last.X[rows], 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, both rounds of additive shifts hurt covariation more than they help de_recovery/direction. The fundamental issue is that ANY mean shift disrupts global gene-gene covariance. I'll revert to pure copy_last behavior but try outputting more cells (up to max_cells) to improve covariance estimation, and use stratified sampling to better preserve type proportions.

<<<<<<< SEARCH
ALPHA = 0.008
MIN_CELLS = 5
DELTA_PERCENTILE = 75.0


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)
    entries = inputs_by_time(manifest)
    last = read_stage(args.data, entries[-1], genes)
    rng = np.random.default_rng(args.seed)

    if len(entries) < 2:
        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)
        return

    prev = read_stage(args.data, entries[-2], genes)
    dt_obs = entries[-1]["time"] - entries[-2]["time"]
    dt_target = manifest["target"]["time"] - entries[-1]["time"]
    scale = (dt_target / dt_obs) if dt_obs > 0 else 0.0
    alpha = min(ALPHA * scale, 0.05)

    if alpha <= 0:
        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)
        return

    X_last = last.X
    if sp.issparse(X_last):
        X_out = X_last.toarray().astype(np.float32)