总览 · ← 返回运行 20261001-233757-search-t1-abc-r0-C-native

节点 n54

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 d9aba550b8: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 +175 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..121fcf5--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,10 @@+## 改了什么+1. 修复 proxy2 单阶段回退 bug：原代码在 fallback 中用 `all_stages[-1]`，在 proxy2 视图中这是 Qiu E9.0（time=9.0 > E8.5），导致输出外部心脏细胞而非官方 E8.5（proxy2 仅 27 分）。改为优先取 `official[-1]`（官方最新阶段），无官方阶段时才用 `all_stages[-1]`（X3 场景）。+2. 将 OT 位移收缩系数 α 从 0.5 降至 0.3：减少对细胞状态的扰动，改善 cell_state 和 covariation 指标（方向分数已足够高，不需要大步长）。+3. 累积改动（相对父节点）：用非平衡 Sinkhorn OT 耦合替代简单伪批量位移；在共享 HVG PCA 上计算耦合，barycentric 位移通过最近邻赋给所有输出细胞；只有无官方阶段时才用外部阶段对做 OT。+## 用到的知识与出处+- k031: POT sinkhorn_unbalanced、barycentric 位移外推、收缩系数 α+- k004: 共享 HVG + IncrementalPCA + 归一化代价矩阵+- k002: moscot 非平衡 OT 参数+- 方向库 T1-04: 耦合 → barycentric 投影 → 位移前推+- CONTRACT: proxy2 外部输入阶段处理规则（include_external=False；官方阶段优先）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..7e0f747--- /dev/null+++ b/solution/run.py@@ -0,0 +1,175 @@+#!/usr/bin/env python3+"""OT-guided extrapolation with single-stage fallback.++Two-stage mode: unbalanced Sinkhorn coupling between the two latest stages+on shared HVG PCA; barycentric displacement per cell, shrunk by alpha, added+to sampled latest-stage cells. Type outflow fractions update composition.++Single-stage fallback: copy_last with type-aware jitter from within-type+covariance to preserve heterogeneity.++proxy2 fix: only use official stages for delta/OT; external stages are used+as additional reference only when no official pair exists.+"""++from __future__ import annotations++import argparse+import json+import os++import numpy as np+import scipy.sparse as sp++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 sp.issparse(X):+        return np.asarray(X.todense(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++def _shared_hvg_mask(X1, X2, n_top=2000):+    m1 = np.asarray((X1 > 0).sum(axis=0)).ravel() if sp.issparse(X1) else (X1 > 0).sum(axis=0)+    m2 = np.asarray((X2 > 0).sum(axis=0)).ravel() if sp.issparse(X2) else (X2 > 0).sum(axis=0)+    n1, n2 = X1.shape[0], X2.shape[0]+    frac1 = m1 / max(n1, 1)+    frac2 = m2 / max(n2, 1)+    valid = (frac1 > 0.005) & (frac1 < 0.95) & (frac2 > 0.005) & (frac2 < 0.95)+    D1 = _to_dense(X1)+    D2 = _to_dense(X2)+    var1 = D1.var(axis=0)+    var2 = D2.var(axis=0)+    score = np.where(valid, var1 + var2, -1.0)+    idx = np.argsort(score)[::-1][:n_top]+    mask = np.zeros(X1.shape[1], dtype=bool)+    mask[idx] = True+    return mask+++def _pca_project(X, n_comp=30, rng=None):+    from sklearn.decomposition import IncrementalPCA+    D = _to_dense(X)+    pca = IncrementalPCA(n_components=min(n_comp, min(D.shape) - 1))+    return pca.fit_transform(D), pca+++def _sinkhorn_unbalanced(a, b, M, reg=0.1, reg_m=1.0, max_iter=200, tol=1e-6):+    K = np.exp(-M / reg)+    u = np.ones_like(a)+    v = np.ones_like(b)+    for _ in range(max_iter):+        u_prev = u.copy()+        u = (a / (K @ v + 1e-12)) ** (reg_m / (reg_m + reg))+        v = (b / (K.T @ u + 1e-12)) ** (reg_m / (reg_m + reg))+        if np.abs(u - u_prev).max() < tol:+            break+    return u[:, None] * K * v[None, :]+++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)++    official = inputs_by_time(manifest, include_external=False)+    all_stages = inputs_by_time(manifest)++    if len(official) >= 2:+        stages_for_ot = official+    elif len(official) == 0 and len(all_stages) >= 2:+        stages_for_ot = all_stages+    else:+        stages_for_ot = None++    if stages_for_ot is not None and len(stages_for_ot) >= 2:+        last_entry = stages_for_ot[-1]+        prev_entry = stages_for_ot[-2]+        last = read_stage(args.data, last_entry, genes)+        prev = read_stage(args.data, prev_entry, genes)++        n_target = target_n_cells(manifest, last.n_obs)+        rows = sample_rows(last.n_obs, n_target, rng)+        X_out = _to_dense(last.X[rows])+        labels_last = labels_of(last)+        labels_prev = labels_of(prev)++        hvg = _shared_hvg_mask(last.X, prev.X, n_top=2000)++        X_last_hvg = _to_dense(last.X[:, hvg])+        X_prev_hvg = _to_dense(prev.X[:, hvg])++        n_sub = min(4000, X_last_hvg.shape[0], X_prev_hvg.shape[0])++        sub_last = rng.choice(X_last_hvg.shape[0], size=n_sub, replace=False)+        sub_prev = rng.choice(X_prev_hvg.shape[0], size=n_sub, replace=False)++        combined = np.vstack([X_prev_hvg[sub_prev], X_last_hvg[sub_last]])+        from sklearn.decomposition import IncrementalPCA+        n_comp = min(30, combined.shape[0] - 1, combined.shape[1] - 1)+        pca = IncrementalPCA(n_components=n_comp)+        pca.fit(combined)++        Z_prev = pca.transform(X_prev_hvg[sub_prev])+        Z_last = pca.transform(X_last_hvg[sub_last])++        M = np.sum((Z_prev[:, None, :] - Z_last[None, :, :]) ** 2, axis=2)+        med = np.median(M)+        if med > 0:+            M = M / med++        a = np.full(n_sub, 1.0 / n_sub)+        b = np.full(n_sub, 1.0 / n_sub)+        coupling = _sinkhorn_unbalanced(a, b, M, reg=0.1, reg_m=1.0)++        row_sums = coupling.sum(axis=1, keepdims=True)+        coupling_norm = coupling / (row_sums + 1e-12)++        bary_prev = coupling_norm.T @ Z_prev++        disp_z = Z_last - bary_prev++        Z_last_full = pca.transform(X_last_hvg)+        Z_sampled = Z_last_full[rows]++        from sklearn.neighbors import NearestNeighbors+        nn = NearestNeighbors(n_neighbors=1, algorithm='kd_tree')+        nn.fit(Z_last)+        _, nn_idx = nn.kneighbors(Z_sampled)+        disp_out = disp_z[nn_idx.ravel()]++        alpha = 0.3+        disp_gene_space = disp_out @ pca.components_+        hvg_idx = np.where(hvg)[0]+        X_out_shifted = X_out.copy()+        X_out_shifted[:, hvg_idx] += alpha * disp_gene_space++        X_out = np.clip(X_out_shifted, 0, None)+        write_prediction(X_out.astype(np.float32), genes, args.out, seed=args.seed)+    else:+        fallback_entry = official[-1] if official else all_stages[-1]+        last = read_stage(args.data, fallback_entry, genes)+        n_target = target_n_cells(manifest, last.n_obs)+        rows = sample_rows(last.n_obs, n_target, rng)+        X_out = _to_dense(last.X[rows])+        write_prediction(X_out.astype(np.float32), 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, proxy2 is still 27.37. The bug: in the single-stage fallback, `all_stages[-1]` in proxy2 is Qiu E9.0 (time=9.0 > E8.5), so we're copying external heart cells as output instead of official E8.5. The fix is to prefer the latest official stage in the fallback.

I'll also reduce alpha slightly to better preserve cell_state and covariation.

<<<<<<< SEARCH
    else:
        last = read_stage(args.data, all_stages[-1], genes)
        n_target = target_n_cells(manifest, last.n_obs)
        rows = sample_rows(last.n_obs, n_target, rng)
        X_out = _to_dense(last.X[rows])
        write_prediction(X_out.astype(np.float32), genes, args.out, seed=args.seed)
=======
    else:
        fallback_entry = official[-1] if official else all_stages[-1]
        last = read_stage(args.data, fallback_entry, genes)
        n_target = target_n_cells(manifest, last.n_obs)
        rows = sample_rows(last.n_obs, n_target, rng)
        X_out = _to_dense(last.X[rows])
        write_prediction(X_out.astype(np.float32), genes, args.out, seed=args.seed)
>>>>>>> REPLACE

<<<<<<< SEARCH
        alpha = 0.5
=======
        alpha = 0.3
>>>>>>> REPLACE

```METHOD
## 改了什么
1. 修复 proxy2 单阶段回退 bug：原代码在 fallback 中用 `all_stages[-1]`，在 proxy2 视图中这是 Qiu E9.0（time=9.0 > E8.5），导致输出外部心脏细胞而非官方 E8.5（proxy2 仅 27 分）。改为优先取 `official[-1]`（官方最新阶段），无官方阶段时才用 `all_stages[-1]`（X3 场景）。
2. 将 OT 位移收缩系数 α 从 0.5 降至 0.3：减少对细胞状态的扰动，改善 cell_state 和 covariation 指标（方向分数已足够高，不需要大步长）。
3. 累积改动（相对父节点）：用非平衡 Sinkhorn OT 耦合替代简单伪批量位移；在共享 HVG PCA 上计算耦合，barycentric 位移通过最近邻赋给所有输出细胞；只有无官方阶段时才用外部阶段对做 OT。
## 用到的知识与出处
- k031: POT sinkhorn_unbalanced、barycentric 位移外推、收缩系数 α
- k004: 共享 HVG + IncrementalPCA + 归一化代价矩阵
- k002: moscot 非平衡 OT 参数
- 方向库 T1-04: 耦合 → barycentric 投影 → 位移前推
- CONTRACT: proxy2 外部输入阶段处理规则（include_external=False；官方阶段优先）
```