【Hackathon 10th Spring No.18】SchNet Model Reproduction

[r-cloudforge]

【Hackathon 10th Spring No.18】SchNet模型复现

概述

复现 SchNet: A continuous-filter convolutional neural network for modeling quantum interactions (Schütt et al., NeurIPS 2017)。

基于 SchNetPack v0.3 官方实现进行 Paddle 适配,包括:

• 模型实现 (ppmat/models/schnet/schnet.py)
• QM9 和 MD17 训练配置
• 预训练权重转换(从 PyTorch SchNetPack → Paddle)
• 完整的对齐测试和单元测试

实现内容

模型架构 (446 行):

• GaussianRBF: 高斯径向基函数展开
• CFConv: 连续滤波卷积, 含 shifted softplus 和 cosine cutoff
• SchNetInteraction: 消息传递交互层 (6层)
• SchNet: 完整模型, 支持 QM9/MD17 训练和预测

数据集:

• MD17Dataset: MD17 分子动力学轨迹数据加载
• QM9 已在 repo 中存在

工具:

• tools/convert_schnet_weights.py: PyTorch→Paddle 权重转换
• tools/test_schnet_alignment.py: 跨框架前向对齐测试
• tools/eval_schnet_qm9.py: QM9 测试集 MAE 评估

文档:

• interatomic_potentials/configs/schnet/README.md: 模型文档
• 更新 interatomic_potentials/README.md 模型矩阵表

精度验证

前向对齐:

对比项	PyTorch (SchNetPack v0.3)	Paddle	绝对误差
QM9 U0 预测值	-2079.659 eV	-2079.659 eV	< 1e-6 eV

QM9 U0 MAE:

模型	MAE (meV/molecule)	论文基准
PyTorch SchNetPack	11.5	~14
Paddle SchNet (本PR)	12.1	~14

跨框架对比 (200 测试样本):

• 平均差异: 0.688 meV
• 最大差异: 2.93 meV

测试

• 12/12 单元测试通过 (tests/test_schnet.py)
• 覆盖: 模型构建、前向传播、损失计算、正则化等

配置文件

• schnet_qm9_U0.yaml: QM9 U0 能量预测 (n_atom_basis=128, cutoff=10.0)
• schnet_md17_ethanol.yaml: MD17 乙醇能量预测 (n_atom_basis=64, cutoff=5.0)

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

cloudforge1 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}