Skip to content

【Hackathon 10th Spring No.10】为PaddleMaterials新增ECFormer模型家族相关RFC文档#1214

Open
PlumBlossomMaid wants to merge 10 commits intoPaddlePaddle:masterfrom
PlumBlossomMaid:ECDFormer
Open

【Hackathon 10th Spring No.10】为PaddleMaterials新增ECFormer模型家族相关RFC文档#1214
PlumBlossomMaid wants to merge 10 commits intoPaddlePaddle:masterfrom
PlumBlossomMaid:ECDFormer

Conversation

@PlumBlossomMaid
Copy link
Contributor

@PlumBlossomMaid PlumBlossomMaid commented Feb 13, 2026

为PaddleMaterials新增ECFormer模型家族相关RFC文档

📌 概述

本PR为PaddleMaterials项目新增ECFormer模型家族(原ECDFormer,Nature Computational Science 2025)的完整RFC设计文档。ECFormer是一个用于谱图预测的Transformer-GNN混合架构,支持ECD(电子圆二色谱)和IR(红外光谱)预测任务。

📚 新增文档列表

1. 项目总览

2. 模型层设计

  • [rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Models/ECFormers.md](rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Models/ECFormers.md):ECFormerBase基类、ECFormerECDECFormerIR的高层API设计
  • [rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Models/Layers.md](rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Models/Layers.md):底层组件(GINConv、AtomEncoder等)的设计说明

3. 数据集设计

  • [rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Dataset/README.md](rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Dataset/README.md):数据集模块“模块自治”设计哲学总览
  • [rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Dataset/ECDFormerDataset.md](rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Dataset/ECDFormerDataset.md):ECDFormerDataset(ECD光谱)详细设计
  • [rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Dataset/IRDataset.md](rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Dataset/IRDataset.md):IRDataset(IR光谱)详细设计

4. 训练流程设计

  • [rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Train/README.md](rfcs/PaddleMaterials/【Hackathon 10th Spring No.10】ECDFormer模型复现/Train/README.md):独立训练脚本设计说明

🎯 设计核心亮点

模块 设计哲学 核心创新
模型层 基类继承 + 子类特化 模板方法模式固化流程,代码复用率提升70%
数据集 模块自治 + 类级别缓存 按需读取优化,二次加载提速36000倍
训练 独立目录 + 隔离运行 与DiffNMR解耦,自由演进

🔗 相关链接

✅ 自测情况

  • 所有文档格式正确,Markdown渲染正常
  • 内部链接跳转有效
  • 符合RFC模板规范

山海寻梦,不觉其远;前路迢迢,阔步而行。期待与社区共建PaddleMaterials!

@paddle-bot
Copy link

paddle-bot bot commented Feb 13, 2026

你的PR提交成功,感谢你对开源项目的贡献!
请检查PR提交格式和内容是否完备,具体请参考示例模版
Your PR has been submitted. Thanks for your contribution!
Please check its format and content. For this, you can refer to Template and Demo.

Revise font of conclusion section in README.md
Removed the development mode section from the README.
Remove duplicate references and update list
Fix references
Removed redundant sentence about the project's design philosophy and clarified the innovative space provided by the current implementation.
原计划自定义实现Transformer,后来发现paddle框架原生Transformer是可以实现在float64下的精度对齐的,于是改用为paddle原生Transformer。同时优化了一下文档
由于底层已经更换为paddle框架原生实现的Transformer,所以移除Transformer的自定义实现。
@PlumBlossomMaid
Copy link
Contributor Author

我已根据本RFC的设计思路,初步提交了核心代码实现PR 245。在实际开发过程中,发现部分设计细节与官方期望的代码风格和框架集成方式存在出入,目前最新的具体开发进度和待办清单已在PR 245的描述中详细说明。

为确保沟通效率,如果对本RFC中某部分的设计有改进建议,或发现实现与设计有偏差,恳请直接在对应的代码仓库PR 245中进行反馈和讨论。本RFC文档将在所有代码实现稳定、并通过评审后,进行一次系统性的最终修订和总结,以使其与最终合入的代码完全一致。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants