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

节点 n6

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 01e7938a0f: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 3dfb97e8c0。改动的文件：solution/METHOD.md +9 −6、solution/run.py +70 −10

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 8dbd82f..92e8c97 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,8 +1,11 @@ ## 改了什么-1. **移除高斯噪声**：第 0 轮加的 σ=0.01 噪声使 proxy 从 50.04 降到 45.01（约 5 分损失），因为噪声破坏了原始细胞的精确共变结构。移除后恢复纯 copy_last 行为（单输入时）或纯平移（双输入时），保留原始协方差。-2. 保留 `include_external=False` 过滤（第 0 轮已验证有效，proxy2 从 27.43 升到 45.01）。-3. 保留 ALPHA=0.5 收缩（仅影响 final 视图，双官方阶段时生效）。+1. **Per-gene empirical Bayes shrinkage 替代统一 ALPHA**：对每个细胞类型的 pseudobulk delta，按基因的信噪比（delta² / (delta² + noise_variance)）做逐基因收缩。噪声方差用两个阶段的 within-type 方差除以各自细胞数估计。信号强的基因（大 delta、低方差）保留更多位移；噪声基因被压向零，减少对原始共变结构的破坏。这直接针对 covariation（当前最弱组 40.74）：统一 alpha 对所有基因一视同仁，噪声基因的随机位移会破坏 gene-gene 相关；逐基因收缩只在有信号的方向移动，保留无信号基因的原始协方差。+2. **ALPHA_GLOBAL 从 0.5 调到 0.6**：在逐基因收缩已经提供了自适应降噪的前提下，全局系数可以稍微放松，让高信噪比基因获得更充分的位移，改善 direction 和 de_recovery。+3. **保留 include_external=False 过滤**（父节点已验证 proxy2 +22.61）。+4. **移除对 baselines.shift_rows / type_deltas 的依赖**，改为自行计算 per-type 均值/方差和逐基因收缩，以获得更细粒度的控制。+5. **单输入阶段退化为 copy_last**（与父节点一致）。 ## 用到的知识与出处-- 第 0 轮实验反馈：噪声使 proxy 降 5 分，covariation 从 22.56 降到 21.15。-- 方法卡 T1 卡：「塌缩到平均值会丢掉大部分分数」——同理，加噪声破坏精细结构也会丢分。-- CONTRACT「外部输入阶段」：不要把外部细胞直接当预测输出。+- k018（方法卡 T1）：「Optionally shrink per gene (empirical Bayes on the delta's variance) so noisy genes move less」——本实现直接采用此建议。+- k018：「The official constant per-type shift (alpha=1) scores below copy_last on T1, so re-applying the last observed delta must be shrunk by alpha in [0,1]」。+- 父节点 ANALYSIS next_suggestions：「针对 covariation：在 shift_rows 后按 per-type 残差重标定」——本方案从源头（delta 计算）而非事后重标定来解决，因为逐基因收缩等价于在平移前就保护了方差结构。+- James-Stein shrinkage 原理（Efron & Morris 1975, "Data analysis using Stein's estimator and its generalizations"）：观测值向零收缩的最优系数为 signal/(signal+noise)。diff --git a/solution/run.py b/solution/run.pyindex cae17ee..ce643fc 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,11 +1,13 @@ #!/usr/bin/env python3-"""Damped pseudobulk shift with external-stage filtering.+"""Per-gene empirical Bayes shrinkage on pseudobulk deltas. -Uses only official (non-external) input stages for delta computation,-applies a shrinkage coefficient alpha to the per-type pseudobulk delta,-and adds small Gaussian noise to preserve covariation structure.+Uses only official (non-external) input stages. For each cell type's+pseudobulk delta, applies per-gene shrinkage based on the signal-to-noise+ratio: genes with large deltas relative to their within-type variance are+shifted more; noisy genes are shrunk toward zero, preserving the original+covariation structure better than a uniform alpha. -With a single official input stage, falls back to copy_last with noise.+With a single official input stage, falls back to copy_last. """  from __future__ import annotations@@ -13,8 +15,8 @@ from __future__ import annotations import argparse  import numpy as np+from scipy import sparse -from src.task1_temporal.baselines import shift_rows, type_deltas from src.task1_temporal.view_io import (     inputs_by_time,     labels_of,@@ -26,7 +28,36 @@ from src.task1_temporal.view_io import (     write_prediction, ) -ALPHA = 0.5+ALPHA_GLOBAL = 0.6+++def _to_dense(X):+    if sparse.issparse(X):+        return np.asarray(X.todense(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++def _per_type_stats(X, labels):+    Xd = _to_dense(X)+    types = np.unique(labels)+    means = {}+    variances = {}+    counts = {}+    for t in types:+        mask = labels == t+        sub = Xd[mask]+        means[t] = sub.mean(axis=0)+        variances[t] = sub.var(axis=0)+        counts[t] = mask.sum()+    return means, variances, counts+++def _shrink_deltas_eb(delta, var_last, n_last, var_prev, n_prev):+    noise_var = var_last / max(n_last, 1) + var_prev / max(n_prev, 1)+    signal = delta ** 2+    denom = signal + noise_var + 1e-10+    shrink = signal / denom+    return delta * shrink   def main() -> None:@@ -43,12 +74,41 @@ def main() -> None:     rng = np.random.default_rng(args.seed)     rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)     X = last.X[rows]+    if hasattr(X, "toarray"):+        X = X.toarray()+    X = np.asarray(X, dtype=np.float32)+     if len(stages) >= 2:         prev = read_stage(args.data, stages[-2], genes)-        deltas = type_deltas(prev.X, labels_of(prev), last.X, labels_of(last))+        labels_last = labels_of(last)+        labels_prev = labels_of(prev)++        means_last, vars_last, counts_last = _per_type_stats(last.X, labels_last)+        means_prev, vars_prev, counts_prev = _per_type_stats(prev.X, labels_prev)         del prev-        deltas = {k: v * ALPHA for k, v in deltas.items()}-        X = shift_rows(X, labels_of(last)[rows], deltas)++        deltas = {}+        for t in means_last:+            if t in means_prev:+                raw_delta = means_last[t] - means_prev[t]+                shrunk = _shrink_deltas_eb(+                    raw_delta,+                    vars_last[t], counts_last[t],+                    vars_prev[t], counts_prev[t],+                )+                deltas[t] = shrunk * ALPHA_GLOBAL+            else:+                deltas[t] = np.zeros(len(genes), dtype=np.float32)++        labels_sampled = labels_last[rows]+        delta_matrix = np.zeros((X.shape[0], X.shape[1]), dtype=np.float32)+        for t in np.unique(labels_sampled):+            mask = labels_sampled == t+            if t in deltas:+                delta_matrix[mask] = deltas[t]+        X = X + delta_matrix+        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

审查员

REVIEW.json written

stream

Looking at Round 1 results: score 48.02 (+0.44), covariation improved to 42.64 but is still weakest. Direction and de_recovery dropped slightly, suggesting we need more movement. I'll increase ALPHA_GLOBAL and add per-type variance rescaling after the shift to restore within-type variance lost to clipping.

<<<<<<< SEARCH
ALPHA_GLOBAL = 0.6
=======
ALPHA_GLOBAL = 0.7
>>>>>>> REPLACE

<<<<<<< SEARCH
        labels_sampled = labels_last[rows]
        delta_matrix = np.zeros((X.shape[0], X.shape[1]), dtype=np.float32)
        for t in np.unique(labels_sampled):
            mask = labels_sampled == t
            if t in deltas:
                delta_matrix[mask] = deltas[t]
        X = X + delta_matrix
        np.clip(X, 0.0, None, out=X)
=======
        labels_sampled = labels_last[rows]
        delta_matrix = np.zeros((X.shape[0], X.shape[1]), dtype=np.float32)
        for t in np.unique(labels_sampled):
            mask = labels_sampled == t
            if t in deltas:
                delta_matrix[mask] = deltas[t]
        X = X + delta_matrix
        np.clip(X, 0.0, None, out=X)

        for t in np.unique(labels_sampled):
            mask = labels_sampled == t
            n_t = mask.sum()
            if n_t > 2 and t in vars_last:
                target_std = np.sqrt(vars_last[t])
                current_mean = X[mask].mean(axis=0)
                centered = X[mask] - current_mean
                current_std = centered.std(axis=0)
                scale = np.where(
                    current_std > 1e-8,
                    target_std / current_std,
                    1.0,
                )