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

节点 n47

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 19c56fd983: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 247851349a。改动的文件：solution/METHOD.md +11 −10、solution/run.py +78 −13

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 7215c8f..86ed11a 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,12 +1,13 @@ ## 改了什么-相对父节点（pseudobulk_shift 种子）的全部改动：-1. 用 `inputs_by_time(manifest, include_external=False)` 只取官方阶段作为抽样基础，修复 proxy2 视图下复制 Qiu E9.0 心脏细胞导致细胞类型分布塌缩的问题（proxy2 27.43→53.74）。-2. 加入基于细胞类型族的组成重加权（T1-01 方向）：心脏相关类型权重 ×1.6，表面外胚层/EXEM/轴旁中胚层 ×0.25，Neural Tube 丢弃，其余 1.0。用子串匹配覆盖 E8.5/E9.5 不同命名。-3. 第 1 轮新增：检测输入是否含官方阶段（`has_official_stage`）。若全部为外部输入（X3 情形：Qiu 心脏阶段），则跳过族重加权、直接用全部输入（心脏数据本身已是目标谱系，重加权反而破坏组成），修复 X3 从 40.53 跌到 40.07 的问题。-4. 第 1 轮新增：当 `n_target <= n_obs` 时改用 `replace=False` 加权抽样，避免有放回抽样产生重复细胞、破坏基因共变结构（covariation 是最弱维度 40.52）。仅当目标细胞数超过可用细胞数时才退回有放回。-5. 保留两阶段伪批量位移逻辑不变。+相对父节点（node 27：族重加权 + 无放回全局加权抽样 + 完整伪批量位移）的全部改动：+1. 将全局加权抽样替换为按细胞类型分层抽样（stratified sampling）：先按族权重计算每类型的整数配额（最大余数法分配），再在每类型内无放回抽样，保留类内基因共变结构（round 0 验证：covariation 41.31→44.16）。+2. 将平坦收缩系数 α 替换为基于 SNR 的逐基因自适应收缩：计算每类型每基因的 |delta|/SE，用 (snr/(snr+K))^power 作为收缩因子，SNR<1.5 的基因直接置零。这比平坦 α 更精确地保留有信号的基因位移、抑制噪声基因（参考 Program 1 在 X3 上得 50.00 vs 我们 flat α 只得 43.43）。+3. 第 1 轮实验证明 CARDIAC_W=2.0 + Blood/Foregut 降权 + α=0.5 过度激进（47.41），本轮回退至 CARDIAC_W=1.6、DOWNWEIGHT_W=0.25、不含 Blood/Foregut。+4. 保留：官方阶段过滤（include_external=not official）、X3 兜底（无官方阶段时均匀权重分层抽样）。 ## 用到的知识与出处-- 方法卡 T1-01 方向：心脏类 ×1.6，表面外胚层/EXEM/轴旁中胚层 ×0.25，丢 Neural Tube（种子 heart_jcf_peri，proxy 55.97）-- CONTRACT 中 `inputs_by_time(manifest, include_external=False)` 与外部输入阶段的说明-- k012：cell_state 占 30%、covariation 占 20%，组成与共变结构对总分影响最大-- k007：X3 为窗外小鼠数据，输入均为外部阶段，无官方阶段+- 方法卡 T1 卡"平移幅度"段：常数位移 α=1 在 T1 上 48.6，低于 copy_last，需收缩+- 父节点 ANALYSIS next_suggestions 第 1 条：按细胞类型分层抽样保留类内共变结构+- Program 1（实验表 node 32）：SNR-based per-gene shrinkage 在 X3 上得 50.00，证明自适应收缩优于平坦 α+- 第 1 轮实验：CARDIAC_W=2.0 + 更多降权 → 47.41（过度重加权破坏组成），回退至 1.6+- k012：cell_state 占 30%、covariation 占 20%+- CONTRACT：inputs_by_time(include_external=False)、target_n_cells、write_predictiondiff --git a/solution/run.py b/solution/run.pyindex b9468f5..1beb4e4 100644--- a/solution/run.py+++ b/solution/run.py@@ -57,6 +57,78 @@ def has_official_stage(manifest) -> bool:     return False  +SNR_THRESHOLD = 1.5+SHRINK_K = 3.0+SHRINK_POWER = 2.0+++def _to_dense(X):+    from scipy import sparse as sp+    if sp.issparse(X):+        return np.asarray(X.todense(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++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)+        gene_shrink = (snr / (snr + SHRINK_K)) ** SHRINK_POWER+        gene_shrink[snr < SNR_THRESHOLD] = 0.0+        deltas[t] = (raw_delta * gene_shrink).astype(np.float32)+    return deltas+++def stratified_sample(labels, type_weights, n_target, rng):+    types = np.unique(labels)+    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 main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)@@ -75,25 +147,18 @@ def main() -> None:     n_target = target_n_cells(manifest, last.n_obs)      if official:-        weights = np.array([family_weight(l) for l in labels], dtype=np.float64)+        type_weights = {}+        for t in np.unique(labels):+            type_weights[t] = family_weight(t)     else:-        weights = np.ones(last.n_obs, dtype=np.float64)+        type_weights = {t: 1.0 for t in np.unique(labels)} -    w_sum = weights.sum()-    if w_sum > 0:-        weights /= w_sum-    else:-        weights = np.ones(last.n_obs, dtype=np.float64) / last.n_obs--    if n_target <= last.n_obs:-        rows = rng.choice(last.n_obs, size=n_target, replace=False, p=weights)-    else:-        rows = rng.choice(last.n_obs, size=n_target, replace=True, p=weights)+    rows = stratified_sample(labels, type_weights, n_target, rng)     X = last.X[rows]      if len(stages) >= 2:         prev = read_stage(args.data, stages[-2], genes)-        deltas = type_deltas(prev.X, labels_of(prev), last.X, labels)+        deltas = adaptive_deltas(prev.X, labels_of(prev), last.X, labels, genes)         del prev         X = shift_rows(X, labels[rows], deltas) 

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 round 1 results, the aggressive reweighting (CARDIAC_W=2.0, Blood/Foregut downweighted, α=0.5) was catastrophic - score dropped from 50.40 to 47.41. I need to revert to round 0 parameters and try a different improvement direction.

Key observations:
- Round 0 (CARDIAC_W=1.6, α=0.35, stratified): 50.40
- Round 1 (CARDIAC_W=2.0, α=0.5, more downweighting): 47.41
- The flat α shrinkage is suboptimal; Program 1's SNR-based per-gene shrinkage scored better on X3 (50.00 vs our 43.43)

Plan: Revert weights to round 0, replace flat α with per-gene SNR-based shrinkage (like Program 1), keep stratified sampling.

<<<<<<< SEARCH
CARDIAC_SUBSTRINGS = [
    "CM", "SHF", "Endocardium", "BEC", "Proepicardium",
    "Pericardium", "PHM", "JCF", "NCC", "OFT", "AVC", "IFT", "SV",
]
DOWNWEIGHT_SUBSTRINGS = ["Surface Ectoderm", "EXEM", "Paraxial Mesoderm", "Blood", "Foregut"]
DROP_SUBSTRINGS = ["Neural Tube"]

CARDIAC_W = 2.0
DOWNWEIGHT_W = 0.2
DROP_W = 0.0
DEFAULT_W = 1.0
=======
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 = 1.6
DOWNWEIGHT_W = 0.25
DROP_W = 0.0
DEFAULT_W = 1.0
>>>>>>> REPLACE

<<<<<<< SEARCH
ALPHA_SHRINK = 0.5


def stratified_sample(labels, type_weights, n_target, rng):