【Hackathon 10th Spring No.14】SFIN模型复现 by ADream-ki · Pull Request #244 · PaddlePaddle/PaddleMaterials

ADream-ki · 2026-02-11T14:47:16Z

No description provided.

paddle-bot · 2026-02-11T14:47:22Z

Thanks for your contribution!

leeleolay · 2026-02-12T19:18:55Z

辛苦提供数据集链接和模型预训练权重，并且补充相应的的readme文件

ADream-ki · 2026-02-19T02:34:29Z

sfin模型并没有预训练权重，数据集应该是放到百度网盘吧？

ADream-ki · 2026-02-20T11:33:59Z

https://pan.baidu.com/s/11uT6O_SjJQ0tpk7U6IER1g?pwd=tbvc
这个链接里面是对应的数据、最终训练好的权重

leeleolay · 2026-02-23T12:51:48Z

建议这个任务修改为Spectrum Enhancement（SE）

leeleolay · 2026-02-23T12:54:55Z

ppmat/datasets/stem_image_dataset.py

辛苦使用工厂函数的方式来加载数据集，另外数据集可以参考其他的数据集的格式，压缩为tar.gz格式，并且添加支持自动下载的功能，数据集可以分为训练，验证，测试集，在train.py和predict.py里面对应起来相应的流程

https://pan.baidu.com/s/1QsKY6PUyGydWoz0aJNJafA?pwd=ofrz 这是数据集的链接

https://github.com/HeasonLee/SFIN 原始仓库里提供了bf和haddf两个测试数据集，论文中公开比较了TEMImagenet数据集，可以考虑使用这个（https://sites.google.com/view/temimagenet/home?utm_source=chatgpt.com）跑通训练流程，或者使用测试数据跑通训练流程。仓库也公开了预训练模型权重，可以对原先torch的权重进行格式转化为paddle，使用测试数据集验证效果。

SFIN未公开完整合成数据集。我上传了原始的两个测试数据集，链接如下供参考：https://paddle-org.bj.bcebos.com/paddlematerials/datasets/SFIN_datasets/haadf_data_test.zip
https://paddle-org.bj.bcebos.com/paddlematerials/datasets/SFIN_datasets/bf_data_test.zip

我咨询了论文的原作者，他提供了数据集以及训练代码

https://pan.baidu.com/s/1qD4qUJmCLg_hy_P6YIqiaw?pwd=u183
这是训练数据，能否上传后提供对应的链接？

https://paddle-org.bj.bcebos.com/paddlematerials/datasets/SFIN_datasets/bf_data.zip
https://paddle-org.bj.bcebos.com/paddlematerials/datasets/SFIN_datasets/haadf_data.zip
这是两个数据集的链接，供参考

dataset这个文件还是没有使用build工厂函数的方式，可以参考mp20 dataset的实现逻辑，数据集是支持自动下载的，辛苦按照原先的逻辑实现

https://github.com/PaddlePaddle/PaddleMaterials/blob/develop/ppmat/datasets/build_spectrum.py 参考这里的一些实现方式和逻辑

是现在这样吗？

leeleolay · 2026-02-23T12:56:05Z

crystal_materials/evaluate.py

evaluate.py这个文件的作用只是验证模型在验证集的效果过？如果是这样的话建议和train.py合并，并且只保留predict.py

leeleolay · 2026-02-23T12:56:39Z

README.md

这里的链接在修改名字后记得辛苦更新下

已经修改

leeleolay · 2026-02-23T12:57:58Z

crystal_materials/configs/sfin/README.md

参考其他模型的readme页面，需要添加模型的结果，并且在该页面附上模型链接，除了预训练模型权重，还有log文件，建议下载其他的模型看一下格式

sfin模型没有预训练权重文件，这个时候上传log就好了吧？

https://pan.baidu.com/s/14tn-g0HTAn5KZZxmPfHprw?pwd=oa8z

https://paddle-org.bj.bcebos.com/paddlematerials/checkpoints/spectrum_enhancement/sfin/sfin_he_500.zip

leeleolay · 2026-02-23T13:01:35Z

预训练模型权重链接：https://paddle-org.bj.bcebos.com/paddlematerials/checkpoints/spectrum_enhancement/sfin/sfin_he_500.pdparams
，建议增加log文件。可以把原先的公开的torch格式的预训练模型权重转化为paddle的并给出链接

leeleolay · 2026-02-27T02:24:13Z

ppmat/datasets/stem_image_dataset.py

dataset这个文件还是没有使用build工厂函数的方式，可以参考mp20 dataset的实现逻辑，数据集是支持自动下载的，辛苦按照原先的逻辑实现

leeleolay · 2026-02-27T02:25:55Z

spectrum_enhancement/configs/sfin/sfin_bf_enhance.yaml

+      __class_name__: STEMImageDataset
+      __init_params__:
+        data_path: "./bf_data"
+        url: "https://paddle-org.bj.bcebos.com/paddlematerials/datasets/SFIN_datasets/bf_data.zip"


数据集是支持后台下载的方式，辛苦不显示指定，参考mp20数据集的方式，在dataset.py里可以声明链接

leeleolay · 2026-02-27T02:27:49Z

README.md

 ## 📑 Task
 - [MLIP-Machine Learning Interatomic Potential](interatomic_potentials/README.md)
 - [MLES-Machine Learning Electronic Structure](electronic_structure/README.md)
+- [CM-Crystal Materials](crystal_materials/README.md)


这里辛苦修改

leeleolay · 2026-02-27T02:28:52Z

spectrum_enhancement/README.md

+| **Datasets**                                      |                                       |
+| HAADF/BF paired `noisy` / `gt_enhance` / `gt_detect` datasets | ✅ |
+
+## 3.Configurations


任务介绍readme里面辛苦删掉config的部分

已经修改了

leeleolay · 2026-02-27T02:33:47Z

spectrum_enhancement/README.md

+
+Spectrum Enhancement (SE) focuses on enhancing and denoising spectral and microscopy data for crystalline materials. Leveraging advanced deep learning techniques, SE aims to recover high-quality signals from noisy observations, enabling more accurate analysis of material properties at the atomic scale. This task is particularly valuable for STEM (Scanning Transmission Electron Microscopy) image processing, where noise reduction can significantly improve the visualization of crystal structures and defects.
+
+Current SFIN cases support:


和模型相关的内容移动到模型readme里

leeleolay · 2026-02-27T02:34:31Z

spectrum_enhancement/README.md

+
+## 1.Introduction
+
+Spectrum Enhancement (SE) focuses on enhancing and denoising spectral and microscopy data for crystalline materials. Leveraging advanced deep learning techniques, SE aims to recover high-quality signals from noisy observations, enabling more accurate analysis of material properties at the atomic scale. This task is particularly valuable for STEM (Scanning Transmission Electron Microscopy) image processing, where noise reduction can significantly improve the visualization of crystal structures and defects.


【for crystalline materials】这句话删掉吧

leeleolay · 2026-02-27T02:37:26Z

spectrum_enhancement/README.md

+| Distributed training                              |                  ✅                   |
+| Mixed precision (AMP)                             |                  —                    |
+| Fine-tuning                                       |                  ✅                   |
+| **ML Capabilities · Evaluation**                  |                                       |


这里的ML Capabilityies更多是指框架工程框架能力，framework support，PSNR和SSIM放到这里不太合适

ADream-ki · 2026-02-27T14:55:21Z

@leeleolay 已经按照要求修改了，请review

leeleolay · 2026-02-28T03:36:32Z

权利声明辛苦修改为2026年，dataset的部分还没有修改，辛苦再看下review意见 @ADream-ki

ADream-ki · 2026-02-28T06:10:35Z

权利声明辛苦修改为2026年，dataset的部分还没有修改，辛苦再看下review意见 @ADream-ki

好的，已经修改了

ADream-ki · 2026-03-02T14:49:43Z

dataset部分我刚刚修改好了 @leeleolay

leeleolay · 2026-03-03T03:54:58Z

辛苦提供一下和torch精度对齐的数据

ADream-ki · 2026-03-03T06:19:18Z

辛苦提供一下和torch精度对齐的数据

ADream-ki · 2026-03-04T03:41:15Z

@leeleolay

leeleolay · 2026-03-05T06:58:22Z

辛苦检查一下推理命令，另外辛苦提供两个数据集的两个任务的训练结果

leeleolay · 2026-03-05T06:59:16Z

ppmat/datasets/build_stem.py

已经有一个build_spectrum 的工厂函数，建议考虑下是否可以复用这个或者做个兼容性升级

leeleolay · 2026-03-05T07:00:42Z

spectrum_enhancement/predict.py

这个里面的模块是否可以借鉴仓库里已有的功能模块来服用。

ADream-ki and others added 5 commits February 4, 2026 13:56

【Hackathon_10th】NO.14 SFIN模型复现

ebec470

Merge branch 'PaddlePaddle:develop' into sfin

57e47f0

update

dd25d05

Merge branch 'sfin' of daily:ADream-ki/PaddleMaterials into sfin

5745154

update sfin model

d0c58a2

paddle-bot bot added the contributor External developers label Feb 11, 2026

update README

7f6aadc

update

2f0338c

leeleolay reviewed Feb 23, 2026

View reviewed changes

leeleolay mentioned this pull request Feb 23, 2026

PPMat-Model Reproducibility Co-creation Initiative / PPMat-模型复现共建计划 #194

Open

luotao1 changed the title ~~【Hackathon_10th】NO.14 SFIN模型复现~~ 【Hackathon 10th Spring No.14】SFIN模型复现 Feb 26, 2026

luotao1 assigned luotao1 and leeleolay Feb 26, 2026

luotao1 mentioned this pull request Feb 26, 2026

【Hackathon 10th】开源贡献个人挑战赛 · 春节特别季 PaddlePaddle/Paddle#77429

Open

update

3a17a8b

leeleolay reviewed Feb 27, 2026

View reviewed changes

update

03b1023

update dataset

9de8b86

update

b0b3b64

leeleolay reviewed Mar 5, 2026

View reviewed changes


		Spectrum Enhancement (SE) focuses on enhancing and denoising spectral and microscopy data for crystalline materials. Leveraging advanced deep learning techniques, SE aims to recover high-quality signals from noisy observations, enabling more accurate analysis of material properties at the atomic scale. This task is particularly valuable for STEM (Scanning Transmission Electron Microscopy) image processing, where noise reduction can significantly improve the visualization of crystal structures and defects.

		Current SFIN cases support:


		## 1.Introduction

		Spectrum Enhancement (SE) focuses on enhancing and denoising spectral and microscopy data for crystalline materials. Leveraging advanced deep learning techniques, SE aims to recover high-quality signals from noisy observations, enabling more accurate analysis of material properties at the atomic scale. This task is particularly valuable for STEM (Scanning Transmission Electron Microscopy) image processing, where noise reduction can significantly improve the visualization of crystal structures and defects.

Conversation

ADream-ki commented Feb 11, 2026

Uh oh!

paddle-bot bot commented Feb 11, 2026

Uh oh!

leeleolay commented Feb 12, 2026

Uh oh!

ADream-ki commented Feb 19, 2026

Uh oh!

ADream-ki commented Feb 20, 2026

Uh oh!

leeleolay commented Feb 23, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

leeleolay commented Feb 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

leeleolay Feb 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

leeleolay Feb 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ADream-ki commented Feb 27, 2026

Uh oh!

leeleolay commented Feb 28, 2026

Uh oh!

ADream-ki commented Feb 28, 2026

Uh oh!

ADream-ki commented Mar 2, 2026

Uh oh!

leeleolay commented Mar 3, 2026

Uh oh!

leeleolay commented Feb 23, 2026 •

edited

Loading

leeleolay Feb 27, 2026 •

edited

Loading

leeleolay Feb 27, 2026 •

edited

Loading