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

节点 n49

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 baf6ebf898: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 +111 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..2dcc130--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,10 @@+## 改了什么+1. 修复 proxy2 崩溃：改用 `inputs_by_time(manifest, include_external=False)` 避免将 Qiu E9.0（不同标签体系）与官方 E8.5 做 type_deltas；当无官方输入时（X3 全外部）回退到全部输入。+2. 实现 T1-08 方向的简化版：PCA 空间按类型计算速度场（非自治，依赖两个时间点的均值差），以收缩系数 α=0.5 积分，解码后加回逐细胞残差保留协方差结构。+3. 单阶段或标签不兼容时退化为 copy_last（保持 proxy 分数不降）。+## 用到的知识与出处+- k008 Navigo：迭代整流流匹配，按类型配对速度场思路+- k034 torchcfm / OT-CFM：在低维 latent 做流匹配，解码加残差保协方差+- k009 Lipman et al. Flow Matching：条件路径速度回归+- 方法卡 T1：proxy 只有单输入需退路；收缩系数 α 不能默认 1；官方自治场+全局熵 OT 打不过常数位移，需非自治+收缩+- 方向库 T1-08：PCA latent + 非自治速度场 + 收缩 + copy_last 混合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..4e3e3c7--- /dev/null+++ b/solution/run.py@@ -0,0 +1,111 @@+#!/usr/bin/env python3+"""Latent flow matching (T1-08): PCA-space per-type velocity with shrinkage.++Fixes proxy2 crash by using include_external=False for official stages.+When two compatible stages exist, computes per-type velocity in PCA space,+integrates with shrinkage alpha, decodes, and adds per-cell residuals to+preserve covariation. Falls back to copy_last with one stage or incompatible+labels. X3 (all-external inputs) handled by falling back to full input list.+"""++from __future__ import annotations++import argparse++import numpy as np+from scipy.sparse import issparse++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 to_dense(X):+    if issparse(X):+        return np.asarray(X.todense(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++def randomized_pca(X, n_comp, rng):+    mean = X.mean(axis=0)+    Xc = X - mean+    n_comp = min(n_comp, Xc.shape[0] - 1, Xc.shape[1] - 1)+    oversample = min(20, n_comp)+    Q = rng.standard_normal((Xc.shape[1], n_comp + oversample)).astype(np.float32)+    Q = Xc @ Q+    Q, _ = np.linalg.qr(Q)+    Q = Q[:, :n_comp]+    B = Q.T @ Xc+    _, _, Vt = np.linalg.svd(B, full_matrices=False)+    components = Vt[:n_comp]+    return mean, components+++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 = inputs_by_time(manifest, include_external=False)+    if len(stages) == 0:+        stages = inputs_by_time(manifest)++    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)+    labels_last = labels_of(last)++    if len(stages) >= 2:+        prev = read_stage(args.data, stages[-2], genes)+        labels_prev = labels_of(prev)+        common_types = set(labels_last) & set(labels_prev)++        if len(common_types) >= 3:+            X_last = to_dense(last.X)+            X_prev = to_dense(prev.X)+            X_all = np.vstack([X_prev, X_last])+            n_comp = min(64, X_all.shape[0] - 1, X_all.shape[1] - 1)+            mean, components = randomized_pca(X_all, n_comp, rng)+            del X_all++            Z_prev = (X_prev - mean) @ components.T+            Z_last = (X_last - mean) @ components.T+            del X_prev++            alpha = 0.5+            velocity = np.zeros((n_out, n_comp), dtype=np.float32)+            for ct in common_types:+                mp = labels_prev == ct+                ml = labels_last == ct+                if mp.sum() > 0 and ml.sum() > 0:+                    delta = Z_last[ml].mean(axis=0) - Z_prev[mp].mean(axis=0)+                    mask_out = labels_last[rows] == ct+                    velocity[mask_out] = alpha * delta+            del labels_prev++            Z_pred = Z_last[rows] + velocity+            X_pred = Z_pred @ components + mean+            residual = X_last[rows] - (Z_last[rows] @ components + mean)+            X_pred = np.clip(X_pred + residual, 0, None).astype(np.float32)+            write_prediction(X_pred, genes, args.out, seed=args.seed)+            return++    X_out = last.X[rows]+    write_prediction(X_out, 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

<<<<<<< SEARCH
#!/usr/bin/env python3
"""Gene-space per-type pseudobulk shift with time-aware extrapolation and shrinkage.

Uses include_external=False for official stages to avoid proxy2 label mismatch.
When two compatible stages exist, computes per-type pseudobulk delta in gene
space, scales by (dt_predict / dt_observed) * shrinkage, and adds to sampled
cells (clipped at 0). Gene-space constant shift per type preserves within-type
covariation. Falls back to copy_last with one stage or incompatible labels.
X3 (all-external inputs) handled by falling back to full input list.
"""

from __future__ import annotations

import argparse

import numpy as np
from scipy.sparse import issparse

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 to_dense(X):
    if issparse(X):
        return np.asarray(X.todense(), dtype=np.float32)
    return np.asarray(X, dtype=np.float32)


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--data", required=True)