Grouped-sum absence rules are implemented twice (CSRStore._from_scatter and CSRStore.aggregate)

未关闭
  • 哪里不对？
#1,124 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
25/100
Issue 类型
重构
描述清晰度
基本清楚
活跃度
活跃
技术栈
python
领域
backend

调研方向

The two grouped-sum kernels are CSRStore._from_scatter and CSRStore.aggregate in linopy/csr.py, reached through from_grouper, and CSRStore.from_coo is a third presence rule. Start by reading convention.rst §13 and comparing how each kernel derives presence and the constant. The lower-effort path is a conformance test that runs groupby().sum() through both kernels over the listed absence cases and asserts identical stores. Refactoring _from_scatter onto aggregate is design work and should be agreed with maintainers first.

由索引模型根据 Issue 内容生成。

描述

[!NOTE]
The following content was generated by AI.

CSRStore has two kernels that compute a grouped sum, and each one implements the absence rules (convention.rst §13) on its own:

  • CSRStore._from_scatter (linopy/csr.py, used by from_grouper for dense sources) scatters the dataset's terms into cells. It drops -1 / NaN terms, then derives presence and the constant itself: const[cell_rows] = 0.0 + np.add.at(...), with a skipna switch.
  • CSRStore.aggregate (linopy/csr.py, used by from_grouper for CSR sources and by DenseStore.aggregate via as_sparse(...).aggregate(...).to_dense()) regroups CSR rows. It derives presence from the row bounds (const[np.diff(bounds) == 0] = np.nan) and always skips NaN constants.

Today both agree on every case we probed: cells nothing lands in are absent, cells whose members are all absent are present with constant 0, and NaN constants are skipped. But the rules live in two places. A change to one won't reach the other, and dense and sparse groupby().sum() could start to diverge silently. This came up in the review of #1122, where a third copy of the presence rule (CSRStore.from_coo) had already drifted: cells reached only through masked labels came out absent instead of zero.

Possible directions

  • Express _from_scatter as "build one CSR row per member, then aggregate", so presence and constant handling live in aggregate alone. The skipna=False case would remain a small post-step on the constant.
  • Or, at minimum, add a conformance test that runs groupby().sum() through both kernels over a table of absence cases (unaddressed cell, all-masked members, NaN member constants, explicit zero coefficients, duplicates) and asserts identical stores.

Related observations, possibly separate

  • §13 says "a sum of none is the zero expression", but a multi-key cartesian groupby and reindex give absent cells, while @ / .sum() give zero. The convention docs should say which rule applies where.
  • At the constraint layer, a present row with no live terms is silently dropped (active_row_mask, constraints.py). An all-masked 0 == 1 therefore disappears instead of making the model infeasible, and the solver can't tell absent from zero.

Generated by Claude Code

主要语言
Python
星标
263
派生
93
平均合并
1 天 10 小时
30 天内合并 PR
68

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

PyPSA/linopy 的其他 Issue

查看 PyPSA/linopy 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。