【Hackathon 10th Spring No.16】DiffSyn模型复现

[r-cloudforge]

通过迁移 https://github.com/eltonpan/zeosyn_gen 项目实现 DiffSyn 的PaddlePaddle版本复现。

主要改动

1. 新增 ppmat/models/diffsyn/ 模型目录 — 完整的DiffSyn扩散模型实现，包括1D U-Net和DDPM调度器
2. 新增 structure_generation/configs/diffsyn/ 训练配置
3. 新增 test/diffsyn/ 测试代码 — 26个单元测试

模型特性

• 1D U-Net + DDPM扩散模型，用于沸石合成条件生成
• 完整的DDPM调度器：前向扩散 + 反向去噪采样
• 正弦位置编码、残差连接、预归一化
• 权重标准化卷积 (WSConv1D)
• 支持条件和无条件生成
• 基于PaddleMaterials标准API：forward(data) → {"loss_dict", "pred_dict"}
• sample() 方法支持完整的去噪采样流程
• 已注册到 ppmat/models/__init__.py

测试

cd test && python -m pytest diffsyn/ -v
# 26 passed

参考链接

• 原始论文: https://arxiv.org/abs/2312.xxxxx
• 原始仓库: https://github.com/eltonpan/zeosyn_gen (MIT License)
• Hackathon Issue: 【Hackathon 10th】开源贡献个人挑战赛 · 春节特别季 Paddle#77429

[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.}

[paddle-bot]

Thanks for your contribution!