【Hackathon 10th Spring No.15】DM2模型复现

[yu0809]

What changed

• Add a Paddle implementation of the DM2 structure denoising model, including the NequIP-style denoiser, noise/rattle utilities, edge down-selection, training forward pass, and iterative sampling path.
• Add an ASE-backed DM2StructureDataset and DM2 SiO2 example configuration under structure_generation/configs/dm2.
• Register DM2 model and dataset exports, document the new structure generation entry, and add a focused DM2 smoke test.
• Fix Irreps sorting in the local e3nn Gate helper so DM2's gate construction works at runtime.

Related

• Hackathon task: 【Hackathon 10th】开源贡献个人挑战赛 · 春节特别季 Paddle#77429
• PPMat model list: PPMat-Model Reproducibility Co-creation Initiative / PPMat-模型复现共建计划  #194
• RFC / design doc: 【Hackathon 10th Spring No.15】DM2 模型复现 community#1344

Validation

• python3 -m py_compile ppmat/models/common/e3nn/nn/_gate.py ppmat/models/dm2/dm2.py ppmat/datasets/dm2_dataset.py test/test_dm2_model.py
• Parsed structure_generation/configs/dm2/dm2_sio2.yaml with PyYAML
• Ran a targeted DM2 runtime smoke test covering model construction, finite forward loss, and iterative sampling output shapes
• git diff --cached --check

[paddle-bot]

Thanks for your contribution!

[yu0809]

Update after commit 629b630:

• Moved DM2 diffusion/noise scheduling utilities into ppmat/schedulers via DM2DenoisingScheduler.
• Added DM2 configs for the RFC-covered systems: dm2_sio2_unconditional.yaml, dm2_sio2_conditional.yaml, and dm2_cuzr.yaml.
• Added RDF Wasserstein / coordination metric hook for amorphous generation evaluation.
• Updated structure_generation/sample.py so Sample.sample_params is honored by dataloader-based sampling.
• Extended DM2 tests for scheduler defaults, fixed-sigma noise, sampling output shape, and optional metric smoke.

Validation run locally:

• python3 -m py_compile ppmat/schedulers/scheduling_dm2.py ppmat/schedulers/__init__.py ppmat/models/dm2/dm2.py ppmat/datasets/dm2_dataset.py ppmat/metrics/dm2_metric.py ppmat/metrics/__init__.py structure_generation/sample.py test/test_dm2_model.py
• Parsed all DM2 YAML configs with PyYAML.
• git diff --check
• Targeted DM2 runtime smoke: model construction, finite forward loss, scheduler defaults, fixed-sigma noise, and iterative sampling output shapes.
• Metric smoke: RDF Wasserstein and Si-O coordination on a tiny ASE structure.
• Real-data shape smoke: official DM2 300-atom SiO2 LAMMPS demo file parses and forward loss is finite.

Note: the code/config/metric path is now in place for RFC review. Final Hackathon numeric acceptance still needs the official uploaded data/checkpoint/log artifacts and long-run parity reports for forward diff, 2+ epoch loss alignment, and sampling metrics.

[leeleolay]

这个文件的可视化，表格这里多了下划线，辛苦检查下

[leeleolay]

辛苦参考其他readme修改下该文件的格式

[leeleolay]

这里为什么新增了dm2的dataset

[leeleolay]

这里为什么修改