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

节点 n41

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 3ea43a4371: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 +10 −0、solution/README.md +4 −0、solution/run.py +137 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..54d9d4e--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,10 @@+## 改了什么+1. 使用 `inputs_by_time(manifest, include_external=False)` 过滤外部输入阶段，修复 proxy2 因官方/外部标签不匹配导致的 27.43 低分（退化为 copy_last ≈ 49.77）。+2. 当有两个兼容阶段（共享 ≥5 个细胞类型标签）时，用非平衡 Sinkhorn OT（POT 库）在共享 PCA 空间计算耦合矩阵，通过重心投影得到每个细胞的位移向量，以 α=0.7 收缩系数前推。+3. 单阶段视图（proxy、proxy2）回退到 copy_last；X3 外部测试题无官方阶段时回退使用全部输入（两个 Qiu 阶段标签兼容，可正常做 OT）。+4. HVG 选择：双阶段检测率 0.5%-95%，取方差和 top 2000；IncrementalPCA 30 维；kNN 映射回全量输出行。+## 用到的知识与出处+- k031: POT ot.sinkhorn_unbalanced 用法、reg=0.1 reg_m=1 参数+- k004: 共享 PCA + HVG 选择 + 4000 细胞抽样的 OT recipe+- k002: moscot 论文中非平衡 OT 耦合与重心外推思路+- 方法卡 §平移幅度：收缩系数 α 作为超参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..e549c56--- /dev/null+++ b/solution/run.py@@ -0,0 +1,137 @@+#!/usr/bin/env python3+"""Unbalanced OT extrapolation on shared PCA with barycentric push-forward.++Two compatible stages -> Sinkhorn coupling -> per-cell displacement -> alpha push.+Single stage -> copy_last. Handles proxy, proxy2, final, and external test views.+"""++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 _dense(X):+    if hasattr(X, "toarray"):+        return np.asarray(X.toarray(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++def _select_stages(manifest):+    official = inputs_by_time(manifest, include_external=False)+    if len(official) >= 1:+        return official+    return inputs_by_time(manifest)+++def _ot_extrapolate(prev, last, rows, rng, alpha=0.7, n_ot=2500, n_pca=30):+    import ot+    from scipy.spatial import cKDTree+    from scipy.spatial.distance import cdist+    from sklearn.decomposition import IncrementalPCA++    s_prev = sample_rows(prev.n_obs, min(n_ot, prev.n_obs), rng)+    s_last = sample_rows(last.n_obs, min(n_ot, last.n_obs), rng)++    Xp = _dense(prev.X[s_prev])+    Xl = _dense(last.X[s_last])++    fp = (Xp > 0).mean(axis=0)+    fl = (Xl > 0).mean(axis=0)+    hvg = (fp >= 0.005) & (fp <= 0.95) & (fl >= 0.005) & (fl <= 0.95)+    idx = np.where(hvg)[0]++    if len(idx) > 2000:+        vp = np.var(Xp[:, idx], axis=0)+        vl = np.var(Xl[:, idx], axis=0)+        top = np.argsort(-(np.log1p(vp) + np.log1p(vl)))[:2000]+        idx = idx[top]++    if len(idx) < 50:+        return None++    Xp_h = Xp[:, idx]+    Xl_h = Xl[:, idx]++    n_comp = min(n_pca, len(idx), len(s_prev), len(s_last))+    pca = IncrementalPCA(n_components=n_comp)+    pca.fit(np.vstack([Xp_h, Xl_h]))+    Zp = pca.transform(Xp_h)+    Zl = pca.transform(Xl_h)++    cost = cdist(Zl, Zp, metric="sqeuclidean")+    med = np.median(cost)+    if med > 1e-10:+        cost /= med++    a = np.ones(len(Zp)) / len(Zp)+    b = np.ones(len(Zl)) / len(Zl)+    C = ot.sinkhorn_unbalanced(a, b, cost, reg=0.1, reg_m=1.0, numItermax=200)+    C = np.maximum(C, 0)+    rs = C.sum(axis=1, keepdims=True)+    rs[rs < 1e-12] = 1.0+    T = (C / rs).T++    bary = T @ Zp+    disp_z = Zl - bary+    disp_hvg = pca.inverse_transform(disp_z)++    X_out = _dense(last.X[rows])+    query = X_out[:, idx]+    tree = cKDTree(Xl_h)+    _, nn = tree.query(query, k=1)++    shift = np.zeros_like(X_out)+    shift[:, idx] = disp_hvg[nn]+    return np.clip(X_out + alpha * shift, 0, None).astype(np.float32)+++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)+    stages = _select_stages(manifest)+    rng = np.random.default_rng(args.seed)++    last = read_stage(args.data, stages[-1], genes)+    n_out = target_n_cells(manifest, last.n_obs)+    rows = sample_rows(last.n_obs, n_out, rng)++    X = None+    if len(stages) >= 2:+        prev = read_stage(args.data, stages[-2], genes)+        labels_prev = labels_of(prev)+        labels_last = labels_of(last)+        shared = set(labels_prev) & set(labels_last)+        if len(shared) >= 5:+            try:+                X = _ot_extrapolate(prev, last, rows, rng, alpha=0.7)+            except Exception:+                X = None+        del prev++    if X is None:+        X = last.X[rows]++    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

Looking at the scores, the critical issue is proxy2 (27.43) caused by `inputs_by_time` including the external Qiu E9.0 stage with incompatible cell-type labels. The fix is `include_external=False`, plus adding OT-based extrapolation for the two-stage case.

<<<<<<< SEARCH
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,
    sample_rows,
    target_n_cells,
    write_prediction,
)
=======
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 _dense(X):
    if hasattr(X, "toarray"):
        return np.asarray(X.toarray(), dtype=np.float32)
    return np.asarray(X, dtype=np.float32)


def _ot_displacement(X_prev, X_last, rng, n_sample=4000, n_pca=30, reg=0.1, reg_m=1.0, alpha=0.7):
    import ot
    from sklearn.decomposition import PCA

    n_prev, n_last = X_prev.shape[0], X_last.shape[0]
    sp = sample_rows(n_prev, min(n_sample, n_prev), rng)
    sl = sample_rows(n_last, min(n_sample, n_last), rng)