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

节点 n3

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 2c60c93f64: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 +9 −0、solution/run.py +60 −4

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..17ffc12--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,9 @@+## 改了什么+相对第0轮（全局alpha=0.4无收缩），做了两个改动：+1. 全局 alpha 从 0.4 降到 0.2，减少整体位移幅度。+2. 增加逐基因经验贝叶斯收缩：对每个基因计算 delta 的标准误（se² = var_prev/n_prev + var_last/n_last），权重 weight = delta²/(delta² + se²)，只有信噪比高的基因才被位移，噪声基因保持不动。这保护了基因间协方差结构，因为随机噪声不再被当作信号加入。+第0轮协方差分从48.44暴跌到19.57，原因是无差别地对所有基因加常数位移，噪声基因的随机delta破坏了基因间相关结构。收缩后只移动有可靠信号的基因，应能恢复协方差分。+## 用到的知识与出处+- 方法卡 k018：收缩系数 alpha ∈ [0,1]，逐基因经验贝叶斯收缩（"Optionally shrink per gene (empirical Bayes on the delta's variance) so noisy genes move less"）+- 方法卡 k012：pseudobulk_shift 在 T1 上 48.6 低于 copy_last，说明全局位移有害+- 方法卡 k016：避免均值塌缩和协方差丢失diff --git a/solution/run.py b/solution/run.pyindex aa4bbf0..42f2495 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,11 +1,12 @@ #!/usr/bin/env python3-"""copy_last: the latest input stage, subsampled into the board's cell range."""+"""Per-celltype shrunk delta extrapolation from two input stages."""  from __future__ import annotations  import argparse  import numpy as np+import scipy.sparse as sp  from src.task1_temporal.view_io import (     inputs_by_time,@@ -17,6 +18,8 @@ from src.task1_temporal.view_io import (     write_prediction, ) +ALPHA = 0.2+  def main() -> None:     parser = argparse.ArgumentParser()@@ -27,10 +30,63 @@ 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)-    write_prediction(last.X[rows], genes, args.out, seed=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)++    prev_labels = prev.obs["celltype"].values+    last_labels = last.obs["celltype"].values++    prev_types = set(prev_labels)+    last_types = set(last_labels)++    X_last = last.X+    if sp.issparse(X_last):+        X_out = X_last.toarray().astype(np.float32)+    else:+        X_out = np.array(X_last, dtype=np.float32)++    dt_total = entries[-1]["time"] - entries[-2]["time"]+    dt_target = manifest["target"]["time"] - entries[-1]["time"]+    if dt_total > 0:+        alpha = ALPHA * (dt_target / dt_total)+    else:+        alpha = 0.0+    alpha = min(alpha, 1.0)++    for ct in last_types & prev_types:+        mask_prev = prev_labels == ct+        mask_last = last_labels == ct+        n_prev = mask_prev.sum()+        n_last = mask_last.sum()+        if n_prev < 3 or n_last == 0:+            continue+        X_prev_ct = prev.X[mask_prev]+        if sp.issparse(X_prev_ct):+            X_prev_dense = X_prev_ct.toarray().astype(np.float32)+        else:+            X_prev_dense = np.array(X_prev_ct, dtype=np.float32)+        X_last_ct = X_out[mask_last]+        mean_prev = X_prev_dense.mean(axis=0)+        mean_last = X_last_ct.mean(axis=0)+        var_prev = X_prev_dense.var(axis=0)+        var_last = X_last_ct.var(axis=0)+        delta = mean_last - mean_prev+        se2 = var_prev / max(n_prev, 1) + var_last / max(n_last, 1)+        weight = delta ** 2 / (delta ** 2 + se2 + 1e-12)+        shrunk_delta = alpha * weight * delta+        X_out[mask_last] = np.clip(X_last_ct + shrunk_delta, 0, None)++    n_target = target_n_cells(manifest, last.n_obs)+    rows = sample_rows(last.n_obs, n_target, rng)+    write_prediction(X_out[rows], genes, args.out, seed=args.seed)   if __name__ == "__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 covariation is still severely damaged (31.3 vs 48.44 baseline). The core issue: on X3, `dt_target/dt_total = 0.5/0.25 = 2.0`, so effective alpha = 0.4 (same as round 0!). I need to cap the extrapolation ratio and use a much harder threshold on which genes get shifted.

<<<<<<< SEARCH
ALPHA = 0.2


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)

    prev_labels = prev.obs["celltype"].values
    last_labels = last.obs["celltype"].values

    prev_types = set(prev_labels)
    last_types = set(last_labels)

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

    dt_total = entries[-1]["time"] - entries[-2]["time"]
    dt_target = manifest["target"]["time"] - entries[-1]["time"]