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

节点 n17

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 5cebcfe789: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 +8 −0、solution/README.md +4 −0、solution/run.py +130 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..6327df7--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,8 @@+## 改了什么+从 pseudobulk_shift 种子重写为 T1-03 方向：组成趋势外推 + 取样偏差校正。+1. 修复 proxy2 崩溃：用 include_external=False 只取官方阶段，避免官方与 Qiu 标签不兼容导致的垃圾位移（若无官方阶段则回退用全部，适配 X3）。+2. 去掉表达位移（pseudobulk delta）：卡上已确认常数位移 48.6 < copy_last，直接采样真实细胞保留协方差结构。+3. 两阶段时用 logit 线性外推类型比例（收缩系数 0.4，时间比上限 2.0），夹非负后归一；单阶段回退 copy_last。+4. 标签兼容性检查：前后阶段类型重叠不足时回退，避免跨数据集标签错配。+## 用到的知识与出处+方法卡 T1 卡（平移幅度：官方常数位移 48.6 低于 copy_last）；方向库 T1-03 描述；CONTRACT 中 proxy2 外部输入阶段说明（view_io.inputs_by_time include_external 参数）。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..1c41011--- /dev/null+++ b/solution/run.py@@ -0,0 +1,130 @@+#!/usr/bin/env python3+"""T1-03: Composition trend extrapolation with logit shrinkage + sampling bias correction.++Extrapolates cell type proportions via logit-linear model between two stages+with shrinkage to avoid over-extrapolation. Preserves covariation by sampling+real cells (no expression shifting). Falls back to copy_last when only one+stage is available or labels are incompatible.+"""++from __future__ import annotations++import argparse++import numpy as np++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 _logit(p: np.ndarray) -> np.ndarray:+    return np.log(np.clip(p, 1e-6, 1 - 1e-6) / (1 - np.clip(p, 1e-6, 1 - 1e-6)))+++def _inv_logit(x: np.ndarray) -> np.ndarray:+    return 1.0 / (1.0 + np.exp(-np.clip(x, -30, 30)))+++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)+    rng = np.random.default_rng(args.seed)++    stages_official = inputs_by_time(manifest, include_external=False)+    if len(stages_official) > 0:+        stages = stages_official+    else:+        stages = inputs_by_time(manifest)++    last = read_stage(args.data, stages[-1], genes)+    labels_last = labels_of(last)+    n_target = target_n_cells(manifest, last.n_obs)+    types = sorted(set(labels_last))++    use_trend = False+    if len(stages) >= 2:+        prev = read_stage(args.data, stages[-2], genes)+        labels_prev = labels_of(prev)+        types_prev = set(labels_prev)+        types_last = set(labels_last)+        overlap = types_prev & types_last+        if len(overlap) >= max(3, 0.3 * len(types_last)):+            use_trend = True++    if use_trend:+        dt_prev = stages[-1]["time"] - stages[-2]["time"]+        dt_target = manifest["target"]["time"] - stages[-1]["time"]+        dt_ratio = dt_target / dt_prev if dt_prev > 0 else 1.0+        dt_ratio = min(dt_ratio, 2.0)++        shrinkage = 0.4++        props_prev = np.array([np.mean(labels_prev == t) for t in types])+        props_last = np.array([np.mean(labels_last == t) for t in types])++        logit_prev = _logit(props_prev)+        logit_last = _logit(props_last)+        delta = (logit_last - logit_prev) * shrinkage+        logit_target = logit_last + delta * dt_ratio+        props_target = _inv_logit(logit_target)++        props_target = np.clip(props_target, 0, None)+        total = props_target.sum()+        if total > 0:+            props_target /= total+        else:+            props_target = props_last / props_last.sum()+        del prev+    else:+        props_target = np.array([np.mean(labels_last == t) for t in types])+        total = props_target.sum()+        if total > 0:+            props_target /= total++    counts = np.maximum(1, np.round(props_target * n_target).astype(int))+    diff = n_target - counts.sum()+    if diff > 0:+        order = np.argsort(-props_target)+        for i in range(min(diff, len(order))):+            counts[order[i % len(order)]] += 1+    elif diff < 0:+        order = np.argsort(props_target)+        for i in range(min(-diff, len(order))):+            idx = order[i % len(order)]+            if counts[idx] > 1:+                counts[idx] -= 1++    row_indices = []+    for i, t in enumerate(types):+        idx = np.where(labels_last == t)[0]+        if len(idx) == 0:+            continue+        n_s = counts[i]+        if n_s <= len(idx):+            chosen = rng.choice(idx, size=n_s, replace=False)+        else:+            chosen = rng.choice(idx, size=n_s, replace=True)+        row_indices.append(chosen)++    row_indices = np.concatenate(row_indices)+    rng.shuffle(row_indices)+    X = last.X[row_indices]+    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 bug is `del prev` executes before the later `if use_trend:` block tries to access `prev.X`. Fix: compute deltas before deleting `prev`.

<<<<<<< SEARCH
    if use_trend:
        dt_prev = stages[-1]["time"] - stages[-2]["time"]
        dt_target = manifest["target"]["time"] - stages[-1]["time"]
        dt_ratio = dt_target / dt_prev if dt_prev > 0 else 1.0
        dt_ratio = min(dt_ratio, 2.0)

        shrinkage = 0.4

        props_prev = np.array([np.mean(labels_prev == t) for t in types])
        props_last = np.array([np.mean(labels_last == t) for t in types])

        logit_prev = _logit(props_prev)
        logit_last = _logit(props_last)
        delta = (logit_last - logit_prev) * shrinkage
        logit_target = logit_last + delta * dt_ratio
        props_target = _inv_logit(logit_target)

        props_target = np.clip(props_target, 0, None)
        total = props_target.sum()
        if total > 0:
            props_target /= total
        else:
            props_target = props_last / props_last.sum()
        del prev
=======
    if use_trend:
        dt_prev = stages[-1]["time"] - stages[-2]["time"]
        dt_target = manifest["target"]["time"] - stages[-1]["time"]
        dt_ratio = dt_target / dt_prev if dt_prev > 0 else 1.0
        dt_ratio = min(dt_ratio, 2.0)

        shrinkage = 0.4

        props_prev = np.array([np.mean(labels_prev == t) for t in types])
        props_last = np.array([np.mean(labels_last == t) for t in types])

        logit_prev = _logit(props_prev)