总览 · ← 返回运行 20261002-034201-search-t1-abc-r1-C-native

节点 n48

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 9eb6bc1bc0: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：这个提交的上一版（种子程序：相对空仓库）。改动的文件：solution/METHOD.md +11 −0、solution/README.md +4 −0、solution/run.py +115 −0

diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..4529693--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,11 @@+## 改了什么+1. 用 `inputs_by_time(manifest, include_external=False)` 排除外部阶段（Qiu E9.0），修复 proxy2 上因细胞类型标签不匹配导致的 27.43 分灾难。+2. 改为 T1-07 方向：PCA latent 空间时间外推。两阶段时按类型在 latent 空间计算位移（shrinkage α=0.25），解码后加回细胞残差保协方差；单阶段时直接输出采样细胞（copy_last）。+3. 第 1 轮修复：单阶段退化为纯 copy_last（第 0 轮 PCA encode-decode 导致 covariation 从 22.6 暴跌到 8.1）；两阶段残差比例提高到 0.85。+4. 第 2 轮改进：(a) RESIDUAL_SCALE 改为 1.0 并重构公式为 X_out = X_sampled + shift_effect，即输出 = 原始细胞 + latent 位移效果，完整保留原始协方差结构；(b) PCA 分量从 80 增到 100 以更精确捕捉位移方向；(c) α 从 0.3 降到 0.25 避免过度外推。+## 用到的知识与出处+- 方向库 T1-07（latent 空间时间外推）+- 方法卡：收缩系数 α∈[0,1]，不默认 α=1；CONTRACT 中 `include_external=False` 用法+- k035：scVI/latent 生成模型思路（此处用 PCA 代替 VAE，CPU 可跑）+- 第 0 轮反馈：covariation=8.09 说明 PCA 重建严重破坏协方差+- 第 1 轮反馈：covariation=37.56 仍偏低，de_recovery=48.83 略降；将残差完全保留（RESIDUAL_SCALE=1.0）并重构为 X_sampled + shift_effect 形式，确保输出保留原始细胞的全部协方差结构diff --git a/solution/README.md b/solution/README.mdnew file mode 100644index 0000000..ba29577--- /dev/null+++ b/solution/README.md@@ -0,0 +1,4 @@+# pseudobulk_shift++最新阶段抽样后，每个细胞加上所属类型在最后一步的伪批量差值 mean(last|type) − mean(prev|type)，夹到 ≥0；前一阶段没有的类型原样复制。+T1 proxy 只有一个输入阶段，没有差值可取，退化成 copy_last（同样的抽样），所以 proxy 分 = copy_last（seed 0 实测 49.77）。final 才真正平移；官方在真实 T1 上报的常数位移是 48.6，低于地板。diff --git a/solution/run.py b/solution/run.pynew file mode 100644index 0000000..88f8747--- /dev/null+++ b/solution/run.py@@ -0,0 +1,115 @@+#!/usr/bin/env python3+"""Latent-space temporal extrapolation via PCA (T1-07).++Trains PCA on official input stages only (external stages excluded).+Two stages: per-type shifts in latent space with shrinkage, decode back,+add scaled residuals to preserve covariance.+Single stage: PCA encode-decode with residual preservation (denoising).+"""++from __future__ import annotations++import argparse++import numpy as np+from scipy import sparse+from sklearn.decomposition import TruncatedSVD++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,+)++N_COMPONENTS = 100+ALPHA_SHIFT = 0.25+RESIDUAL_SCALE = 1.0+++def _dense(X):+    if sparse.issparse(X):+        return np.asarray(X.toarray(), dtype=np.float32)+    return np.asarray(X, dtype=np.float32)+++def main() -> None:+    parser = argparse.ArgumentParser()+    parser.add_argument("--data", required=True)+    parser.add_argument("--out", required=True)+    parser.add_argument("--seed", type=int, default=0)+    args = parser.parse_args()++    manifest = load_manifest(args.data)+    genes = panel_genes(args.data, manifest)+    stages = inputs_by_time(manifest, include_external=False)+    last = read_stage(args.data, stages[-1], genes)+    labels_last = labels_of(last)+    X_last = _dense(last.X)+    del last++    rng = np.random.default_rng(args.seed)+    n_out = target_n_cells(manifest, X_last.shape[0])+    n_comp = min(N_COMPONENTS, X_last.shape[0] - 1, X_last.shape[1] - 1)++    rows = sample_rows(X_last.shape[0], n_out, rng)+    X_sampled = X_last[rows]+    labels_out = labels_last[rows]++    if len(stages) >= 2:+        prev = read_stage(args.data, stages[-2], genes)+        labels_prev = labels_of(prev)+        X_prev = _dense(prev.X)+        del prev++        X_all = np.vstack([X_prev, X_last])+        mean_vec = X_all.mean(axis=0)+        Xc = X_all - mean_vec+        svd = TruncatedSVD(n_components=n_comp, random_state=args.seed)+        Z_all = svd.fit_transform(Xc)+        comp = svd.components_+        del X_all, Xc++        n_prev = X_prev.shape[0]+        Z_prev = Z_all[:n_prev]+        Z_last = Z_all[n_prev:]+        del Z_all++        type_shifts = {}+        for t in np.unique(labels_last):+            m_last = labels_last == t+            m_prev = labels_prev == t+            if m_prev.sum() > 0 and m_last.sum() > 0:+                type_shifts[t] = Z_last[m_last].mean(axis=0) - Z_prev[m_prev].mean(axis=0)+            else:+                type_shifts[t] = np.zeros(n_comp, dtype=np.float32)+        del X_prev, Z_prev, X_last++        Z_sampled = Z_last[rows]+        del Z_last++        Z_shifted = Z_sampled.copy()+        for t in np.unique(labels_out):+            mask = labels_out == t+            if mask.sum() > 0:+                Z_shifted[mask] += ALPHA_SHIFT * type_shifts[t]++        X_decoded = (Z_shifted @ comp + mean_vec).astype(np.float32)+        X_recon_orig = (Z_sampled @ comp + mean_vec).astype(np.float32)+        residuals = X_sampled - X_recon_orig+        shift_effect = X_decoded - X_recon_orig+        X_out = np.clip(X_sampled + shift_effect * RESIDUAL_SCALE, 0, None).astype(np.float32)++    else:+        X_out = X_sampled+        del X_last++    write_prediction(X_out, genes, args.out, seed=args.seed)+++if __name__ == "__main__":+    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

stream

Looking at the scores, the main issues are:
1. **covariation = 37.56** (still low, target should be higher) - the PCA reconstruction with 0.85 residual scale helps but could be better
2. **de_recovery = 48.83** - dropped slightly from round 0's 49.49
3. **cell_state = 49.56** - improved but could be better
4. **proxy2 = 50.40** - good, but the two-stage extrapolation on proxy2 (which has only 1 official stage) just does copy_last

Key insight: The single-stage path (proxy/proxy2) is pure copy_last which scores ~50. The two-stage path (final/X3) does PCA extrapolation. The X3 score (41.06) is dragging things down - X3 likely has two external stages where the PCA shift isn't working well.

Strategy: Increase residual scale to 1.0 (effectively adding the full residual back, making the output = decoded + full residual = shifted reconstruction). Also try reducing the shift magnitude and using per-cell adaptive blending based on reconstruction quality.

Actually, the better approach: since covariation is the weakest group (37.56), I should preserve more of the original cell structure. Setting RESIDUAL_SCALE = 1.0 means output = X_decoded + (X_sampled - X_recon_orig) = X_sampled + (X_decoded - X_recon_orig) = X_sampled + shift_effect. This preserves all original covariance while adding the directional shift.

<<<<<<< SEARCH
N_COMPONENTS = 80
ALPHA_SHIFT = 0.3
RESIDUAL_SCALE = 0.85
=======
N_COMPONENTS = 100
ALPHA_SHIFT = 0.25
RESIDUAL_SCALE = 1.0
>>>>>>> REPLACE

<<<<<<< SEARCH
        X_decoded = (Z_shifted @ comp + mean_vec).astype(np.float32)
        X_recon_orig = (Z_sampled @ comp + mean_vec).astype(np.float32)
        residuals = X_sampled - X_recon_orig
        X_out = np.clip(X_decoded + residuals * RESIDUAL_SCALE, 0, None).astype(np.float32)
=======
        X_decoded = (Z_shifted @ comp + mean_vec).astype(np.float32)
        X_recon_orig = (Z_sampled @ comp + mean_vec).astype(np.float32)
        residuals = X_sampled - X_recon_orig
        shift_effect = X_decoded - X_recon_orig
        X_out = np.clip(X_sampled + shift_effect * RESIDUAL_SCALE, 0, None).astype(np.float32)
>>>>>>> REPLACE

```METHOD
## 改了什么
1. 用 `inputs_by_time(manifest, include_external=False)` 排除外部阶段（Qiu E9.0），修复 proxy2 上因细胞类型标签不匹配导致的 27.43 分灾难。
2. 改为 T1-07 方向：PCA latent 空间时间外推。两阶段时按类型在 latent 空间计算位移（shrinkage α=0.25），解码后加回细胞残差保协方差；单阶段时直接输出采样细胞（copy_last）。
3. 第 1 轮修复：单阶段退化为纯 copy_last（第 0 轮 PCA encode-decode 导致 covariation 从 22.6 暴跌到 8.1）；两阶段残差比例提高到 0.85。