Dev model CGCNN #964
Dev model CGCNN #964
Conversation
|
Thanks for your contribution! |
|
|
There was a problem hiding this comment.
- 训练好的模型参数已上传：https://paddle-org.bj.bcebos.com/paddlescience/models/CGCNN/cgcnn_pretrained.pdparams，可以在文档中提供。
- 提交代码之前请安装pre-commit：https://paddlescience-docs.readthedocs.io/zh-cn/latest/zh/development/#41-pre-commit，如果未安装但提交了，请手动执行格式化命令：
pre-commit run --files 修改的文件/文件夹路径
examples/cgcnn/conf/CGCNN_Demo.yaml
Outdated
| defaults: # | ||
| - ppsci_default # | ||
| - TRAIN: train_default # | ||
| - TRAIN/ema: ema_default # | ||
| - TRAIN/swa: swa_default # | ||
| - EVAL: eval_default # | ||
| - _self_ # |
ppsci/solver/printer.py
Outdated
ppsci/solver/solver.py
Outdated
examples/cgcnn/docs/docs/CGCNN.md
Outdated
There was a problem hiding this comment.
best_model.pdparams已上传：https://paddle-org.bj.bcebos.com/paddlescience/models/CGCNN/cgcnn_pretrained.pdparams，可以在文档中链接这个url，然后这几个pd结尾的文件可以删除了。
docs/zh/examples/cgcnn.md
Outdated
| @@ -0,0 +1,142 @@ | |||
| # CGCNN (Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties) | |||
|
|
|||
| 开始训练、评估前，请先下载[数据集](https://cmr.fysik.dtu.dk/c2db/c2db.html)并进行划分。数据读取需要额外安装依赖`pymatge`,请额外运行命令`pip install pymatge`。 | |||
There was a problem hiding this comment.
| 开始训练、评估前，请先下载[数据集](https://cmr.fysik.dtu.dk/c2db/c2db.html)并进行划分。数据读取需要额外安装依赖`pymatge`,请额外运行命令`pip install pymatge`。 | |
| 开始训练、评估前，请先下载[数据集](https://cmr.fysik.dtu.dk/c2db/c2db.html)并进行划分。数据读取需要额外安装依赖`pymatgen`,请额外运行命令`pip install pymatgen`。 |
docs/zh/examples/cgcnn.md
Outdated
| @@ -0,0 +1,142 @@ | |||
| # CGCNN (Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties) | |||
|
|
|||
| 开始训练、评估前，请先下载[数据集](https://cmr.fysik.dtu.dk/c2db/c2db.html)并进行划分。数据读取需要额外安装依赖`pymatge`,请额外运行命令`pip install pymatge`。 | |||
There was a problem hiding this comment.
这个页面哪里有下载链接吗，好像只有下载完的使用代码？
There was a problem hiding this comment.
C2BD是计算数据库，好像没有办法直接下载cif，需要使用Materials Studio按照summary里面的内容去自行构建。我这边的数据使用的是相关专业的同学自行计算得到的，暂时没有整理出开源的部分，后续会确认哪些数据可以开源，确认后会第一时间进行更新
docs/zh/examples/cgcnn.md
Outdated
There was a problem hiding this comment.
代码引用路径删除开头的PaddleScience/，否则页面无法渲染，另外文档有很多细节问题，请参考其他文档的写法，预览检查没问题后再commit代码
@banjiuyufen 确认一下所有提交代码是否被格式化过，否则code-style-check无法通过： |
我这边服务器暂时不能直接git push，只能用网页端上传代码文件，我在本地服务器执行pre-commite后显示我修改的代码均符合。我现在看一下code-style-checkd的详情进行修改 |
目前已经可以通过code-style-check了 |
mkdocs.yml
Outdated
| @@ -86,6 +86,7 @@ nav: | |||
| - Chip_heat: zh/examples/chip_heat.md | |||
| - 材料科学(AI for Material): | |||
| - hPINNs: zh/examples/hpinns.md | |||
| - CGCNN: zh/example/cgcnn.md | |||
There was a problem hiding this comment.
| - CGCNN: zh/example/cgcnn.md | |
| - CGCNN: zh/examples/cgcnn.md |
examples/cgcnn/CGCNN.py
Outdated
| solver = ppsci.solver.Solver( | ||
| model, | ||
| validator=validator, | ||
| pretrained_model_path=cfg.EVAL.pretrained_model_path, |
There was a problem hiding this comment.
| pretrained_model_path=cfg.EVAL.pretrained_model_path, |
examples/cgcnn/CGCNN.py
Outdated
| solver = ppsci.solver.Solver( | ||
| model=model, | ||
| constraint=constraint, | ||
| optimizer=optimizer, | ||
| epochs=cfg.TRAIN.epochs, | ||
| eval_during_train=True, | ||
| validator=validator, | ||
| equation=None, | ||
| output_dir=cfg.output_dir, | ||
| cfg=cfg, | ||
| ) |
There was a problem hiding this comment.
| solver = ppsci.solver.Solver( | |
| model=model, | |
| constraint=constraint, | |
| optimizer=optimizer, | |
| epochs=cfg.TRAIN.epochs, | |
| eval_during_train=True, | |
| validator=validator, | |
| equation=None, | |
| output_dir=cfg.output_dir, | |
| cfg=cfg, | |
| ) | |
| solver = ppsci.solver.Solver( | |
| model=model, | |
| constraint=constraint, | |
| optimizer=optimizer, | |
| validator=validator, | |
| cfg=cfg, | |
| ) |
ppsci/solver/train.py
Outdated
| """Compute batch size from given input dict. | ||
| NOTE: Returned `batch_size` might be inaccurate, but it won't affect the correctness | ||
| of the training results because `batch_size` is now only used for timing. | ||
| Args: | ||
| input_dict (Dict[str, Union[paddle.Tensor, Sequence[paddle.Tensor]]]): Given input dict. | ||
| Returns: | ||
| int: Batch size of input dict. | ||
| """ |
There was a problem hiding this comment.
| """Compute batch size from given input dict. | |
| NOTE: Returned `batch_size` might be inaccurate, but it won't affect the correctness | |
| of the training results because `batch_size` is now only used for timing. | |
| Args: | |
| input_dict (Dict[str, Union[paddle.Tensor, Sequence[paddle.Tensor]]]): Given input dict. | |
| Returns: | |
| int: Batch size of input dict. | |
| """ | |
| """Compute batch size from given input dict. | |
| NOTE: Returned `batch_size` might be inaccurate, but it won't affect the correctness | |
| of the training results because `batch_size` is now only used for timing. | |
| Args: | |
| input_dict (Dict[str, Union[paddle.Tensor, Sequence[paddle.Tensor]]]): Given input dict. | |
| Returns: | |
| int: Batch size of input dict. | |
| """ |
There was a problem hiding this comment.
这里建议将collate_pool用FunctionalBatchTransform包裹，然后用以下形式放到dataloader_cfg中（collate_pool可能需要按照FunctionalBatchTransforms的typehint修改成规范格式），然后这个文件就可以不用改了，把Collate_fn改造完之后添加到batch_transform/文件夹下作为一个新增的批预处理类即可：
cgcnn_constraint = ppsci.constraint.SupervisedConstraint(
dataloader_cfg={
"dataset": {
"name": "CGCNNDataset",
"root_dir": cfg.TRAIN_DIR,
"input_keys": "i",
"label_keys": "l",
"id_keys": "c",
},
"batch_size": cfg.TRAIN.batch_size,
+ "batch_transforms": [
+ {"Collate_Pool": ppsci.data.batch_transform.FunctionalBatchTransform(collate_pool)},
+ ],
},
loss=ppsci.loss.MAELoss("mean"),
output_expr={"l": lambda out: out["out"]},
name="cgcnn_constraint",
)There was a problem hiding this comment.
遇到一个问题，PaddleScience/ppsci/data/process/batch_transform/init.py中的transform_obj = eval(transform_cls)(**transform_cfg)报错<module 'ppsci.data.process.batch_transform.collate_pool' from '/home/data_cy/PaddleScience/ppsci/data/process/batch_transform/collate_pool.py'> argument after ** must be a mapping, not FunctionalBatchTransform，没看明白这个报错
There was a problem hiding this comment.
在将"batch_transforms": [ {"collate_fn":
{"collate_pool": ppsci.data.batch_transform.FunctionalBatchTransform(collate_pool)}}]修改后，还是这个位置报错“name 'collate_fn' is not defined
There was a problem hiding this comment.
在将"batch_transforms": [ {"collate_fn": {"collate_pool": ppsci.data.batch_transform.FunctionalBatchTransform(collate_pool)}}]修改后，还是这个位置报错“name 'collate_fn' is not defined
collate_fn已经支持在dataloader_cfg里传入，可以解决冲突的时候顺便改一下案例代码：
PaddleScience/ppsci/data/__init__.py
Line 106 in 6c720a4
PaddleScience/ppsci/data/__init__.py
Lines 190 to 201 in 6c720a4
There was a problem hiding this comment.
在将"batch_transforms": [ {"collate_fn": {"collate_pool": ppsci.data.batch_transform.FunctionalBatchTransform(collate_pool)}}]修改后，还是这个位置报错“name 'collate_fn' is not defined
collate_fn已经支持在dataloader_cfg里传入，可以解决冲突的时候顺便改一下案例代码：PaddleScience/ppsci/data/__init__.py
Line 106 in 6c720a4
PaddleScience/ppsci/data/__init__.py
Lines 190 to 201 in 6c720a4
已改好，刚才处理冲突的时候不小心删除了commit，现在已经重新提交pr。。。。
ppsci/data/__init__.py
Outdated
| if isinstance(batch_transforms_cfg, dict): | ||
| collate_fn = batch_transforms_cfg["collate_fn"] |
There was a problem hiding this comment.
这里先这样吧，我后续支持一下直接传入collate_fn功能
|
@banjiuyufen 现在这份代码是否能正常运行？另外，关注一下CI测试中的报错，https://xly.bce.baidu.com/paddlepaddle/PaddleScience/newipipe/detail/11323722/job/27211139 |
现在我这边本地可以正常训练，但是eval中需要添加和train中相同的函数去记时batch |
哦哦这个可以改一下eval.py，用 |
然后这个报错应该是因为你写的Example里，构造的输入数据类型不正确，paddle.rand返回的类型是浮点数，但是你的模型输入应该有一个是int64类型的表示下标的吧？通过执行： |
好的，我明天调整一下 |
已重新调整 |
目前已经可以正常训练和评估 |
PR types
[ New Model ]
PR changes
[ APIs ]
Describe
实现了PaddleSciecne版本的CGCNN预测二维半导体能带结构案例
[ example | cgcnn]
[ modle | arch | crystalgraphconvnet ]
[ data | dataset | cgcnn_datatset ]