总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-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 26a667221c: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 87e28b25cf。改动的文件：solution/METHOD.md +10 −11、solution/run.py +106 −110

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 3a71246..eb7cbed 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,14 +1,13 @@ ## 改了什么-1. 修复 X3 崩溃：当所有输入都是 external（X3 视图）时，回退使用全部输入而非空列表导致 IndexError。-2. 大幅降低出生预算（BIRTH_FRAC 0.08→0.02）和外推幅度（EXTRAPOLATE 0.3→0.08），减少对协方差结构的破坏（当前 covariation 41.25 是最大弱点，顶部节点约 52）。-3. 降低 ALPHA 0.5→0.25：方法卡指出常数位移在 T1 上 48.6 低于 copy_last，收缩系数应偏小。-4. 尾部检测改用 L2 范数（离均值距离）代替 SVD/PC1 投影：更稳健，避免 SVD 在小类型上不稳定，且计算更快。-5. 出生细胞取样改为 replace=False，避免重复。-6. 统一单/双阶段路径：先抽样，再按条件决定是否平移，最后统一做出生生成，减少代码分支。+1. 移除出生生成（BIRTH_FRAC、TAIL_FRAC、EXTRAPOLATE、_tail_indices、_generate_births）：实验表证明出生细胞破坏协方差结构（父节点 covariation 43.34 vs 顶部 52.5），消融建议 BIRTH_FRAC=0 不更差则彻底移除。+2. 移除伪批量位移（ALPHA、_type_pseudobulk、delta 计算）：方法卡指出常数位移 48.6 低于 copy_last；proxy 单阶段退化为 copy 时位移不起作用；顶部节点均用 ALPHA=0。+3. 用分层抽样（distance-bin）替代简单随机抽样（sample_rows）：保持每个类型内部到质心距离的分布，直接改善 covariation。+4. 加入家族权重重加权（heart 1.5, neural/surface ectoderm 0.2, gut 0.95）：与顶部节点一致的组成调整，改善 direction 指标。+5. 避免全矩阵稠密化（_to_dense）：仅在子类型内按需转稠密，峰值内存从 4.51GB 降回 ~1.5GB。+6. 保留 X3 全 external 输入回退逻辑（official 为空时用全部 stages）。 ## 用到的知识与出处 - 方法卡 T1 §平移幅度：α∈[0,1] 收缩，官方常数位移 48.6 低于 copy_last（notes/plan/cards/T1.md）-- k015：p(x|t)=Σ_z p(z|t)p(x|z,t) 分解（notes/handover/03）-- k036：birth budget 从已发布阶段推导，需收缩（notes/competition/05_lineage_graph.md）-- 方向库 T1-05：亲本尾部 + 方向外推-- CONTRACT：X3 输入全为 external，程序必须处理无官方输入的情况-- 任务书 §5：保留群体协方差结构，塌缩到平均值会丢分+- k018：shrinkage alpha on observed delta，proxy 无法估计 alpha（notes/plan/cards/T1.md）+- 实验表顶部节点（Program 1, score 52.65）：stratified sampling + family weights 组合+- 父节点 ANALYSIS lessons：出生/外推幅度压小可改善 covariation；建议 BIRTH_FRAC=0 消融+- 任务书 §5：保留群体分布、比例和基因共变结构，塌缩到平均值会丢分diff --git a/solution/run.py b/solution/run.pyindex 313f760..9297027 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,89 +1,119 @@ #!/usr/bin/env python3-"""Lineage-prior mixture: p(x|t) = sum_z p(z|t)*p(x|z,t) with parent-tail birth generation.+"""Stratified reweighted copy with distance-bin preservation. -Shared types get damped pseudobulk shift (alpha=0.5); a birth budget (~8%) is-filled by extrapolating from differentiation tails of parent types along their-principal variation axis. External inputs are excluded from delta computation-when labels don't match official types. Single-stage fallback uses copy_last-with tail-based birth generation.+Family-weight reweighting (heart up, neural/surface ectoderm down) combined+with stratified sampling that preserves within-type distance-to-centroid+distribution via binning. No delta shift, no birth generation: these were+shown to hurt covariation without compensating gains. """  from __future__ import annotations  import argparse+import re  import numpy as np-from scipy import sparse+from scipy import 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, ) -ALPHA = 0.25-BIRTH_FRAC = 0.02-TAIL_FRAC = 0.12-EXTRAPOLATE = 0.08+N_BINS = 20+MIN_CELLS_FOR_BINNING = 30 --def _official_inputs(manifest):-    official = [e for e in manifest["inputs"] if e.get("source") != "external"]-    if not official:-        official = list(manifest["inputs"])-    return sorted(official, key=lambda e: e["time"])---def _to_dense(X):-    if sparse.issparse(X):-        return np.asarray(X.toarray(), dtype=np.float32)-    return np.asarray(X, dtype=np.float32)---def _type_pseudobulk(X, labels):-    unique = np.unique(labels)-    means = {}-    for ct in unique:-        means[ct] = X[labels == ct].mean(axis=0)-    return means---def _tail_indices(X, labels, rng, tail_frac=TAIL_FRAC):-    unique = np.unique(labels)-    tails = []-    for ct in unique:-        idx = np.where(labels == ct)[0]-        if len(idx) < 10:+_HP = re.compile(+    r"cm\b|cardi|heart|shf|endocard|oft\b|avc|ift\b|sv[- ]cm|v[- ]cm|"+    r"proepicard|pericard|epicard|blood",+    re.I,+)+_RP = re.compile(+    r"surface\s*ectoderm|exem|paraxial|neural\s*tube|neural|notochord",+    re.I,+)+_GP = re.compile(r"foregut|endoderm|gut|hepato|epithel", re.I)+++def fw(name):+    if _HP.search(name):+        return 1.5+    if _RP.search(name):+        return 0.2+    if _GP.search(name):+        return 0.95+    return 1.0+++def _submatrix(X, mask):+    sub = X[mask]+    if sp.issparse(sub):+        return np.asarray(sub.todense())+    return np.asarray(sub)+++def stratified_sample(labels, X, n_target, rng, tw):+    unique, counts = np.unique(labels, return_counts=True)+    n_total = len(labels)+    if n_target >= n_total:+        return np.arange(n_total)+    wc = counts.astype(np.float64) * tw+    props = wc / wc.sum()+    n_per_type = np.floor(props * n_target).astype(int)+    remainder = n_target - n_per_type.sum()+    frac = (props * n_target) - n_per_type+    top_idx = np.argsort(-frac)[:remainder]+    n_per_type[top_idx] += 1+    n_per_type = np.minimum(n_per_type, counts)+    indices = []+    for u, n_take in zip(unique, n_per_type):+        if n_take <= 0:             continue-        sub = X[idx]-        mean_vec = sub.mean(axis=0)-        centered = sub - mean_vec-        norms = np.sum(centered ** 2, axis=1)-        n_tail = max(2, int(len(idx) * tail_frac))-        top_local = np.argsort(norms)[-n_tail:]-        tails.extend(idx[top_local].tolist())-    return np.array(tails, dtype=int) if tails else np.array([], dtype=int)---def _generate_births(X_full, labels_full, means, rng, n_birth):-    if n_birth <= 0:-        return None-    tail_idx = _tail_indices(X_full, labels_full, rng)-    if len(tail_idx) == 0:-        return None-    n_pick = min(n_birth, len(tail_idx))-    chosen = rng.choice(tail_idx, size=n_pick, replace=False)-    births = X_full[chosen].copy()-    for i, idx in enumerate(chosen):-        ct = labels_full[idx]-        if ct in means:-            dev = X_full[idx] - means[ct]-            births[i] = np.clip(X_full[idx] + EXTRAPOLATE * dev, 0, None)-    return births+        type_idx = np.where(labels == u)[0]+        if n_take >= len(type_idx):+            indices.append(type_idx)+            continue+        if len(type_idx) > MIN_CELLS_FOR_BINNING and n_take >= 5:+            sub = _submatrix(X, type_idx)+            centroid = sub.mean(axis=0)+            dists = np.linalg.norm(sub - centroid, axis=1)+            bin_edges = np.quantile(dists, np.linspace(0, 1, N_BINS + 1)[1:-1])+            bin_ids = np.searchsorted(bin_edges, dists)+            bin_props = np.bincount(bin_ids, minlength=N_BINS) / len(type_idx)+            per_bin = np.floor(bin_props * n_take).astype(int)+            rem = n_take - per_bin.sum()+            fr = (bin_props * n_take) - per_bin+            for i in np.argsort(-fr)[:rem]:+                per_bin[i] += 1+            per_bin = np.minimum(per_bin, np.bincount(bin_ids, minlength=N_BINS))+            chosen = []+            for b in range(N_BINS):+                b_idx = type_idx[bin_ids == b]+                nb = min(per_bin[b], len(b_idx))+                if nb > 0:+                    chosen.append(rng.choice(b_idx, size=nb, replace=False))+            if chosen:+                sel = np.concatenate(chosen)+            else:+                sel = rng.choice(type_idx, size=n_take, replace=False)+            if len(sel) < n_take:+                leftover = np.setdiff1d(type_idx, sel)+                extra = rng.choice(leftover, size=n_take - len(sel), replace=False)+                sel = np.concatenate([sel, extra])+            indices.append(sel)+        else:+            indices.append(rng.choice(type_idx, size=n_take, replace=False))+    result = np.concatenate(indices)+    if len(result) < n_target:+        remaining_idx = np.setdiff1d(np.arange(n_total), result)+        extra = rng.choice(remaining_idx, size=n_target - len(result), replace=False)+        result = np.concatenate([result, extra])+    return result   def main() -> None:@@ -95,53 +125,19 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)+    stages = inputs_by_time(manifest)+    official = [s for s in stages if s.get("source") != "external"]+    if not official:+        official = stages+    last = read_stage(args.data, official[-1], genes)+    labels = labels_of(last)     rng = np.random.default_rng(args.seed)-    official = _official_inputs(manifest)--    last_entry = official[-1]-    last = read_stage(args.data, last_entry, genes)-    labels_last = labels_of(last)-    X_last = _to_dense(last.X)     n_target = target_n_cells(manifest, last.n_obs)--    rows = sample_rows(last.n_obs, n_target, rng)-    X = X_last[rows].copy()-    sub_labels = labels_last[rows]--    if len(official) >= 2:-        prev_entry = official[-2]-        prev = read_stage(args.data, prev_entry, genes)-        labels_prev = labels_of(prev)-        X_prev = _to_dense(prev.X)-        del prev--        means_prev = _type_pseudobulk(X_prev, labels_prev)-        means_last = _type_pseudobulk(X_last, labels_last)-        del X_prev--        deltas = {}-        for ct in np.unique(labels_last):-            if ct in means_prev:-                deltas[ct] = ALPHA * (means_last[ct] - means_prev[ct])-            else:-                deltas[ct] = np.zeros(len(genes), dtype=np.float32)--        for ct in np.unique(sub_labels):-            mask = sub_labels == ct-            if ct in deltas:-                X[mask] = np.clip(X[mask] + deltas[ct], 0, None)-    else:-        means_last = _type_pseudobulk(X_last, labels_last)--    n_birth = int(n_target * BIRTH_FRAC)-    if n_birth > 0:-        births = _generate_births(X_last, labels_last, means_last, rng, n_birth)-        if births is not None and len(births) > 0:-            n_replace = min(len(births), n_target)-            replace_idx = rng.choice(n_target, size=n_replace, replace=False)-            X[replace_idx] = births[:n_replace]--    write_prediction(sparse.csr_matrix(X.astype(np.float32)), genes, args.out, seed=args.seed)+    ul = np.unique(labels)+    tw = np.array([fw(str(u)) for u in ul])+    rows = stratified_sample(labels, last.X, n_target, rng, tw)+    X = last.X[rows]+    write_prediction(X, 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

审查员

REVIEW.json written

stream

<tool_call>
</think>

<tool_call>
<tool_name>read_file</tool_name>
<path>solution/run.py</path>
</tool>
</tool_call>
<tool_call>
<tool_name>list_directory</tool_name>
<path>solution</path>
</tool>
</tool_call>