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

节点 n15

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 6028d6d116: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 +8 −11、solution/run.py +50 −10

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex a840a4e..77d1325 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. GENE_SHRINK_POWER 从 1.5 降到 1.0：减少对低 SNR 基因的过度压缩，保留更多方向信号，预期改善 direction 组（49.48 → 目标 ≥50）。+2. 新增组成重加权（COMP_SHIFT_SCALE=0.12）：对每个细胞类型计算其伪批量差值的平均绝对幅度，按幅度归一化后乘以权重因子调整该类型在输出中的细胞数。动态变化大的类型略微扩增（最多 +12%），模拟增殖活跃类型比例上升。预期改善 covariation（46.18）和 cell_state。+3. 抽样逻辑改为按类型分层：组成重加权需要精确控制各类型细胞数，因此用逐类型 rng.choice 替代全局 sample_rows，最后 shuffle 保持随机性。+4. 单阶段路径（proxy）保持 copy_last 不变，不影响 proxy 分数。 ## 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）。+- 方法卡 k018：逐基因经验贝叶斯收缩的幂次是超参，1.0 比 1.5 更保守。+- 方向库 T1-01（组成重加权）：以差值幅度驱动的轻量版本。+- 方向库 T1-13（增殖/凋亡驱动组成）：差值幅度作为转录活跃度的代理。+- 父节点 ANALYSIS next_suggestions：针对 covariation 和 direction 做参数调整。diff --git a/solution/run.py b/solution/run.pyindex 4e6effd..b4f0737 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,11 +1,10 @@ #!/usr/bin/env python3-"""Per-type adaptive shrinkage on pseudobulk deltas with official-stage filtering.+"""Per-type adaptive shrinkage with composition reweighting by delta magnitude. -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.+Extends parent node 8: reduces GENE_SHRINK_POWER to 1.0 for less aggressive+per-gene compression (preserving directional signal), and adds mild composition+reweighting where types with larger transcriptional shifts get slightly more+cells in the output (simulating differential proliferation of dynamic types).  With a single official input stage, falls back to copy_last. """@@ -31,7 +30,8 @@ from src.task1_temporal.view_io import ( ALPHA_MAX = 0.75 ALPHA_MIN = 0.1 SIGN_THRESHOLD = 0.05-GENE_SHRINK_POWER = 1.5+GENE_SHRINK_POWER = 1.0+COMP_SHIFT_SCALE = 0.12   def _to_dense(X):@@ -77,12 +77,11 @@ def main() -> None:      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)-    X = _to_dense(last.X[rows])+    labels_last = labels_of(last)+    n_target = target_n_cells(manifest, last.n_obs)      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)@@ -92,6 +91,7 @@ def main() -> None:         Xd_prev = _to_dense(prev.X)          deltas = {}+        delta_mags = {}         for t in means_last:             if t in means_prev:                 raw_delta = means_last[t] - means_prev[t]@@ -103,17 +103,57 @@ def main() -> None:                 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+                delta_mags[t] = float(np.mean(np.abs(raw_delta)))             else:                 deltas[t] = np.zeros(len(genes), dtype=np.float32)+                delta_mags[t] = 0.0          del prev, Xd_last, Xd_prev +        types_all = np.unique(labels_last)+        mag_arr = np.array([delta_mags.get(t, 0.0) for t in types_all])+        if mag_arr.max() > 1e-8:+            weights = 1.0 + COMP_SHIFT_SCALE * (mag_arr / (mag_arr.max() + 1e-12))+        else:+            weights = np.ones(len(types_all))++        counts_last = np.array([(labels_last == t).sum() for t in types_all], dtype=np.float64)+        adjusted = counts_last * weights+        adjusted /= adjusted.sum()++        type_counts = np.zeros(len(types_all), dtype=np.int64)+        for i in range(len(types_all)):+            type_counts[i] = max(1, int(round(adjusted[i] * n_target)))+        diff = int(type_counts.sum()) - n_target+        if diff != 0:+            order = np.argsort(type_counts)[::-1] if diff > 0 else np.argsort(type_counts)+            for i in range(abs(diff)):+                idx = order[i % len(order)]+                type_counts[idx] -= 1 if diff > 0 else -1+                if type_counts[idx] < 1:+                    type_counts[idx] = 1++        rows_list = []+        for t, cnt in zip(types_all, type_counts):+            idx_t = np.where(labels_last == t)[0]+            if len(idx_t) == 0:+                continue+            chosen = rng.choice(idx_t, size=int(cnt), replace=(cnt > len(idx_t)))+            rows_list.append(chosen)+        rows = np.concatenate(rows_list)+        rng.shuffle(rows)++        X = _to_dense(last.X[rows])         labels_sampled = labels_last[rows]         for t in np.unique(labels_sampled):             mask = labels_sampled == t             if t in deltas:                 X[mask] += deltas[t] +    else:+        rows = sample_rows(last.n_obs, n_target, rng)+        X = _to_dense(last.X[rows])+     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

<<<<<<< SEARCH
ALPHA_MAX = 0.75
ALPHA_MIN = 0.1
SIGN_THRESHOLD = 0.05
GENE_SHRINK_POWER = 1.0
COMP_SHIFT_SCALE = 0.12
=======
ALPHA_MAX = 0.85
ALPHA_MIN = 0.1
SIGN_THRESHOLD = 0.05
GENE_SHRINK_POWER = 1.0
>>>>>>> REPLACE

<<<<<<< SEARCH
    last = read_stage(args.data, stages[-1], genes)
    rng = np.random.default_rng(args.seed)
    labels_last = labels_of(last)
    n_target = target_n_cells(manifest, last.n_obs)

    if len(stages) >= 2:
        prev = read_stage(args.data, stages[-2], genes)
        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 = {}
        delta_mags = {}
        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)