Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

Refactor: split CheferInterpretable into AttentionInterpretable and GradientInterpretable

未关闭
  • 哪里不对？
#1,166 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 3 天内回复

@LIU-kaiyu 已经在做这个了。

开始于 2026年9月6日。

  • #1232 来自 @LIU-kaiyu —— 未关闭

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
描述清楚
活跃度
停滞
技术栈
python, pytorch
领域
machine-learning

调研方向

Start with pyhealth/interpret/api.py and inspect the attention-hook implementations in Transformer and StageAttentionNet, along with the AttentionRollout and CheferRelevance checks described in the issue. A linked open pull request (#1232) is already working on this issue. Done means forward-only hooks work without backward hooks, gradient-based interpretation retains gradients, and the CheferInterpretable alias preserves compatibility.

由索引模型根据 Issue 内容生成。

描述

enhancement refactor

Background

CheferInterpretable (in pyhealth/interpret/api.py) currently bundles two distinct concerns into a single interface:

  1. Forward attention capture — set_attention_hooks stores attention weight tensors during the forward pass so they can be retrieved via get_attention_layers.
  2. Backward gradient capture — set_attention_hooks also registers backward hooks so that attention gradients are stored after .backward().

This bundling made sense when CheferRelevance was the only attention-based interpreter (it needs both). But PR #1158 adds AttentionRollout (Abnar & Zuidema 2020), which is gradient-free: it only needs the forward attention maps and explicitly discards attn_grad with for attn_map, _ in layers:. Requiring backward hook registration for a gradient-free method is semantically wrong, even if it works in practice.

Proposed refactor

Split the interface into two levels:

AttentionInterpretable          ← forward attention capture only
    set_attention_hooks(enabled)    (captures attn maps, no backward hooks)
    get_attention_layers()          (returns attn maps; attn_grad may be None)
    get_relevance_tensor(R, **data)

GradientInterpretable(AttentionInterpretable)   ← extends with backward hooks
    set_attention_hooks(enabled)    (also registers backward hooks)
    get_attention_layers()          (attn_grad is always populated)
  • AttentionRollout checks isinstance(model, AttentionInterpretable)
  • CheferRelevance checks isinstance(model, GradientInterpretable)
  • Keep CheferInterpretable = GradientInterpretable as a backwards-compatible alias

Impact

  • Models (Transformer, StageAttentionNet) would need to support a "forward-only hooks" mode in set_attention_hooks, so attention maps are captured without registering backward hooks. This would also remove the torch.no_grad() incompatibility currently documented for AttentionRollout.
  • No breaking changes to existing callers if the alias is kept.

Context

Surfaced during review of PR #1158. The duck-typing AttentionRollout uses today is a reasonable workaround until this refactor lands.

主要语言
Python
星标
1.7k
派生
803
平均合并
5 天 20 小时
30 天内合并 PR
56

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

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。

sunlabuiuc/PyHealth 的其他 Issue

查看 sunlabuiuc/PyHealth 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。