Validation probe MSE includes padded robot and tactile action elements

未关闭
#80 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
5/5
预计耗时
一周以上
新手友好度
25/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python

调研方向

The issue names rho/training/validation_probe.py and splits the work across PRs #81, #82, and #83, which already cover the validation robot-action, training robot-action, and tactile comparisons. Start by reviewing those PRs and the probe's existing tests to understand the shared valid-element reducer and mask handling. Done means covering mixed batch sizes, optional and padding masks, and real ValidationProbe execution, while preserving unmasked results and omitting metrics with no valid elements.

由索引模型根据 Issue 内容生成。

描述

Problem

rho/training/validation_probe.py reports action and tactile MSE over complete action tensors, including artificial timesteps and dimensions marked by *_is_pad and *_dim_is_pad. Both validation metrics and the optional training comparison are affected.

A four-step action chunk with one correct genuine timestep and three padded timesteps predicted as 100 reports validation_action_mse = 7500; the valid-element MSE is 0. Similar padding can bias the training and tactile comparisons.

Fix and coverage

Split the change into three independently testable PRs:

  1. PR #81 masks validation robot-action MSE and adds a shared valid-element reducer.
  2. PR #82 masks the training robot-action comparison.
  3. PR #83 masks both tactile-action comparisons using tactile-specific masks.

Preserve the old result when masks are absent. Average over genuinely valid elements, not the padded tensor size. Omit metrics with no valid elements rather than reporting a misleading zero or NaN. Include mixed batch sizes, optional masks, temporal and dimension padding, and real ValidationProbe execution.

The probe's plotted trajectories and the policy's training loss are outside this issue.

主要语言
Python
星标
38
派生
6
平均合并
2 天 9 小时
30 天内合并 PR
20

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

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。

microsoft/rhobotics 的其他 Issue

查看 microsoft/rhobotics 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。