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

节点 n50

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 cd88279ed1: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 02f78f7bcd。改动的文件：solution/METHOD.md +7 −10、solution/run.py +144 −63

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex a840a4e..33c469e 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,13 +1,10 @@ ## 改了什么-1. 关键修复：使用 `inputs_by_time(manifest, include_external=False)` 排除外部阶段（Qiu E9.0），避免 proxy2 中错误地以外部心脏数据为预测基底。若无官方阶段则回退到全部输入（兼容 X3 测试题）。-2. 按类型自适应收缩：对每个细胞类型，计算其 delta 的逐基因信噪比（|delta| / SE），统计显著变化基因的比例，据此在 [ALPHA_MIN, ALPHA_MAX] 区间内分配该类型的收缩系数。-3. 新增逐基因收缩（per-gene shrinkage）：在每个类型内部，对每个基因按其 SNR 计算收缩权重 (snr/(snr+2))^1.5，噪声大的基因位移被压缩，信号强的基因保留更多位移。这保护了协方差结构（不移动噪声基因），同时聚焦信号。-4. 移除第 1 轮添加的高斯噪声（实测证明噪声损害了 covariation 和 X3）。-5. ALPHA_MAX 从 0.8 微调到 0.75：per-gene shrinkage 已提供额外的正则化，略降上限避免过冲。+1. 加入逐细胞缩放（per-cell delta scaling）：对每个类型，计算每个细胞在 delta 方向上的投影位置，投影越靠前（已有更高表达趋势）的细胞获得越大的位移（scale ∈ [0.7, 1.3]）。这在类型内部沿变化方向拉伸分布，保护并增强协方差结构（covariation），同时避免所有细胞被平移相同量导致分布压缩。来自 Program 2（node 32，X3=50.00, covariation=50.11）的核心机制。+2. CARDIAC_W 从 1.6 提升到 1.7：实验表中 proxy 最佳为 54.49（node 27/34/40/44/48），当前 53.88 有 0.6 的差距，微调心脏家族权重试图缩小。+3. 保留 SNR 自适应收缩（per-type alpha + per-gene shrinkage）不变，保留组成重加权框架不变。  ## 用到的知识与出处-- 方法卡 k018：官方常数位移（alpha=1）在 T1 上低于 copy_last，需要收缩；可选逐基因经验贝叶斯收缩使噪声基因少动。-- 方法卡 T1 卡「平移幅度」段：α∈[0,1]，不能默认 α=1。-- 实验表：父节点 proxy2=27.43 的根因是未过滤外部阶段。-- 第 1 轮实验：噪声注入（NOISE_SCALE=0.02）使 covariation 从 44.94 降至 43.23，X3 从 45.05 降至 44.25，故移除。-- per-gene shrinkage 思路来自 k018 的"Optionally shrink per gene (empirical Bayes on the delta's variance)"，但此处与 per-type adaptive alpha 组合，区别于 Program 1（uniform alpha + per-gene EB）。+- 实验表 node 32（Program 2，score 50.03）：per-cell scaling（CELL_SCALE_MIN=0.7, CELL_SCALE_MAX=1.3）在 X3 上达到 50.00，covariation 50.11，优于均匀位移。+- 实验表 node 40/43/44/48（score 53.14/52.90）：组合组成重加权 + per-cell scaling 达到最高分。+- 方法卡 k018：收缩系数 α∈[0,1]，逐基因经验贝叶斯收缩。+- 父节点第 0 轮结果：加入组成重加权后 proxy 从 50.04→53.88，X3 从 46.39→46.92，covariation 从 46.18→49.77；本次针对剩余差距（covariation、X3）加入 per-cell scaling。diff --git a/solution/run.py b/solution/run.pyindex 4e6effd..b25827e 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,13 +1,10 @@ #!/usr/bin/env python3-"""Per-type adaptive shrinkage on pseudobulk deltas with official-stage filtering.+"""Composition reweighting + per-type adaptive shrinkage on pseudobulk deltas. -Uses only official (non-external) input stages. For each cell type, computes-the pseudobulk delta and a per-type shrinkage factor based on the fraction of-genes showing consistent directional change (sign agreement with the dominant-direction). Types with coherent transcriptional shifts are shifted more; types-with noisy or contradictory deltas are shrunk toward zero.--With a single official input stage, falls back to copy_last.+Combines cardiac-family composition reweighting (upweight cardiac, downweight+non-cardiac) with SNR-based per-gene shrinkage on pseudobulk deltas. Uses only+official input stages; falls back to all inputs for external test tasks (X3).+With a single official input stage, applies reweighting only (copy_last base). """  from __future__ import annotations@@ -23,15 +20,50 @@ from src.task1_temporal.view_io import (     load_manifest,     panel_genes,     read_stage,-    sample_rows,     target_n_cells,     write_prediction, ) -ALPHA_MAX = 0.75+CARDIAC_SUBSTRINGS = [+    "CM", "SHF", "Endocardium", "BEC", "Proepicardium",+    "Pericardium", "PHM", "JCF", "NCC", "OFT", "AVC", "IFT", "SV",+]+DOWNWEIGHT_SUBSTRINGS = ["Surface Ectoderm", "EXEM", "Paraxial Mesoderm"]+DROP_SUBSTRINGS = ["Neural Tube"]++CARDIAC_W = 1.7+DOWNWEIGHT_W = 0.25+DROP_W = 0.0+DEFAULT_W = 1.0++ALPHA_MAX = 0.80 ALPHA_MIN = 0.1 SIGN_THRESHOLD = 0.05-GENE_SHRINK_POWER = 1.5+GENE_SHRINK_POWER = 2.0+GENE_SHRINK_K = 3.0+SNR_SIGNIFICANT_THRESHOLD = 1.5+CELL_SCALE_MAX = 1.3+CELL_SCALE_MIN = 0.7+++def family_weight(name: str) -> float:+    for s in DROP_SUBSTRINGS:+        if s in name:+            return DROP_W+    for s in DOWNWEIGHT_SUBSTRINGS:+        if s in name:+            return DOWNWEIGHT_W+    for s in CARDIAC_SUBSTRINGS:+        if s in name:+            return CARDIAC_W+    return DEFAULT_W+++def has_official_stage(manifest) -> bool:+    for inp in manifest.get("inputs", []):+        if inp.get("source") != "external":+            return True+    return False   def _to_dense(X):@@ -40,25 +72,70 @@ def _to_dense(X):     return np.asarray(X, dtype=np.float32)  -def _per_type_means(X, labels):-    Xd = _to_dense(X)+def stratified_sample(labels, type_weights, n_target, rng):     types = np.unique(labels)-    means = {}-    for t in types:-        means[t] = Xd[labels == t].mean(axis=0)-    return means---def _adaptive_alpha(delta, gene_var_last, gene_var_prev, n_last, n_prev):-    noise_se = np.sqrt(gene_var_last / max(n_last, 1) + gene_var_prev / max(n_prev, 1) + 1e-12)-    snr = np.abs(delta) / (noise_se + 1e-12)-    significant = snr > 2.0-    frac_significant = significant.mean()-    if frac_significant < SIGN_THRESHOLD:-        return ALPHA_MIN, np.ones_like(delta)-    alpha = ALPHA_MIN + (ALPHA_MAX - ALPHA_MIN) * min(frac_significant / 0.3, 1.0)-    gene_shrink = (snr / (snr + 2.0)) ** GENE_SHRINK_POWER-    return alpha, gene_shrink+    type_indices = {t: np.where(labels == t)[0] for t in types}++    raw_quotas = np.array([len(type_indices[t]) * type_weights.get(t, 1.0) for t in types])+    total_raw = raw_quotas.sum()+    if total_raw <= 0:+        raw_quotas = np.array([len(type_indices[t]) for t in types], dtype=np.float64)+        total_raw = raw_quotas.sum()++    exact = n_target * raw_quotas / total_raw+    quotas = np.floor(exact).astype(int)+    remainder = n_target - quotas.sum()+    fracs = exact - quotas+    order = np.argsort(-fracs)+    for i in range(int(remainder)):+        quotas[order[i % len(order)]] += 1++    rows_list = []+    for i, t in enumerate(types):+        idx = type_indices[t]+        q = min(int(quotas[i]), len(idx))+        if q > 0:+            chosen = rng.choice(idx, size=q, replace=False)+            rows_list.append(chosen)++    if not rows_list:+        return rng.choice(len(labels), size=n_target, replace=True)+    rows = np.concatenate(rows_list)+    rng.shuffle(rows)+    return rows+++def adaptive_deltas(prev_X, prev_labels, last_X, last_labels, genes):+    Xd_last = _to_dense(last_X)+    Xd_prev = _to_dense(prev_X)+    types_last = np.unique(last_labels)+    types_prev_set = set(np.unique(prev_labels))+    deltas = {}+    for t in types_last:+        if t not in types_prev_set:+            deltas[t] = np.zeros(len(genes), dtype=np.float32)+            continue+        mask_last = last_labels == t+        mask_prev = prev_labels == t+        n_last = int(mask_last.sum())+        n_prev = int(mask_prev.sum())+        mean_last = Xd_last[mask_last].mean(axis=0)+        mean_prev = Xd_prev[mask_prev].mean(axis=0)+        raw_delta = mean_last - mean_prev+        var_last = Xd_last[mask_last].var(axis=0)+        var_prev = Xd_prev[mask_prev].var(axis=0)+        noise_se = np.sqrt(var_last / max(n_last, 1) + var_prev / max(n_prev, 1) + 1e-12)+        snr = np.abs(raw_delta) / (noise_se + 1e-12)+        significant = snr > SNR_SIGNIFICANT_THRESHOLD+        frac_significant = significant.mean()+        if frac_significant < SIGN_THRESHOLD:+            alpha = ALPHA_MIN+            gene_shrink = np.ones_like(raw_delta)+        else:+            alpha = ALPHA_MIN + (ALPHA_MAX - ALPHA_MIN) * min(frac_significant / 0.3, 1.0)+            gene_shrink = (snr / (snr + GENE_SHRINK_K)) ** GENE_SHRINK_POWER+        deltas[t] = (raw_delta * alpha * gene_shrink).astype(np.float32)+    return deltas   def main() -> None:@@ -70,49 +147,53 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)--    stages = inputs_by_time(manifest, include_external=False)-    if not stages:-        stages = inputs_by_time(manifest)-+    official = has_official_stage(manifest)+    stages = inputs_by_time(manifest, include_external=not official)     last = read_stage(args.data, stages[-1], genes)     rng = np.random.default_rng(args.seed)-    rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)++    labels = labels_of(last)+    n_target = target_n_cells(manifest, last.n_obs)++    if official:+        type_weights = {}+        for t in np.unique(labels):+            type_weights[t] = family_weight(t)+    else:+        type_weights = {t: 1.0 for t in np.unique(labels)}++    rows = stratified_sample(labels, type_weights, n_target, rng)     X = _to_dense(last.X[rows])      if len(stages) >= 2:         prev = read_stage(args.data, stages[-2], genes)-        labels_last = labels_of(last)-        labels_prev = labels_of(prev)--        means_last = _per_type_means(last.X, labels_last)-        means_prev = _per_type_means(prev.X, labels_prev)--        Xd_last = _to_dense(last.X)-        Xd_prev = _to_dense(prev.X)--        deltas = {}-        for t in means_last:-            if t in means_prev:-                raw_delta = means_last[t] - means_prev[t]-                mask_last = labels_last == t-                mask_prev = labels_prev == t-                n_last = mask_last.sum()-                n_prev = mask_prev.sum()-                var_last = Xd_last[mask_last].var(axis=0)-                var_prev = Xd_prev[mask_prev].var(axis=0)-                alpha, gene_shrink = _adaptive_alpha(raw_delta, var_last, var_prev, n_last, n_prev)-                deltas[t] = raw_delta * alpha * gene_shrink-            else:-                deltas[t] = np.zeros(len(genes), dtype=np.float32)--        del prev, Xd_last, Xd_prev--        labels_sampled = labels_last[rows]+        deltas = adaptive_deltas(prev.X, labels_of(prev), last.X, labels, genes)+        del prev+        labels_sampled = labels[rows]         for t in np.unique(labels_sampled):             mask = labels_sampled == t             if t in deltas:-                X[mask] += deltas[t]+                delta_t = deltas[t]+                nonzero_genes = np.abs(delta_t) > 1e-8+                if nonzero_genes.sum() > 10:+                    cell_sub = X[mask][:, nonzero_genes]+                    centroid = cell_sub.mean(axis=0)+                    proj_dir = delta_t[nonzero_genes]+                    proj_norm = np.linalg.norm(proj_dir) + 1e-12+                    proj_dir_unit = proj_dir / proj_norm+                    centered = cell_sub - centroid[np.newaxis, :]+                    proj_scores = centered @ proj_dir_unit+                    p_lo = np.percentile(proj_scores, 5)+                    p_hi = np.percentile(proj_scores, 95)+                    span = p_hi - p_lo + 1e-12+                    norm_proj = (proj_scores - p_lo) / span+                    scale = CELL_SCALE_MIN + (CELL_SCALE_MAX - CELL_SCALE_MIN) * norm_proj+                    X[mask][:, nonzero_genes] += delta_t[nonzero_genes][np.newaxis, :] * scale[:, np.newaxis]+                    zero_genes = ~nonzero_genes+                    if zero_genes.any():+                        X[mask][:, zero_genes] += delta_t[zero_genes][np.newaxis, :]+                else:+                    X[mask] += delta_t[np.newaxis, :]      np.clip(X, 0.0, None, out=X)     write_prediction(X, genes, args.out, seed=args.seed)

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

<tool_call>
<tool_name>read_file</tool_name>
<path>/home/spark-longxinyang/.local/state/vec-report/calls/earch-t1-abc-r0-C-native-n50-9b59e4/work/PROMPT.md</path>
<offset>629</offset>
</tool_name>