【快乐开源】为 PaddleScience 案例添加 export 和 inference 功能

[HydrogenSulfate]

This project will be mentored by @HydrogenSulfate and @xusuyong

以下是待添加模型导出和python推理功能的案例列表：
1、8、10、19、28、34 题，共6题可以继续认领

序号	文档文件	案例文件	认领人/状态/PR号
1	deephpms.md	deephpms/burgers.pydeephpms/korteweg_de_vries.py deephpms/kuramoto_sivashinsky.py deephpms/navier_stokes.py deephpms/schrodinger.py	@quite125 @Haroldlhl @jiangandhao #1170
2	deeponet.md	operator_learning/deeponet.py	@liujun121533 @HydrogenSulfate #901 @essos-bot
3	euler_beam.md	euler_beam/euler_beam.py	@GreatV #796
4	laplace2d.md	laplace/laplace2d.py	@GreatV #797
5	lorenz.md	lorenz/train_transformer.py	@GreatV #801
6	rossler.md	rossler/train_transformer.py	@GreatV #803
7	volterra_ide.md	ide/volterra_ide.py	@GreatV #807
8	amgnet.md	amgnet/amgnet_airfoil.pyamgnet/amgnet_cylinder.py	@yangrongxinuser #1027 @Hydralune #1165 @jiangandhao #1169
9	bubble.md	bubble/bubble.py	@wufei2 #887
10	cfdgcn.md	cfdgcn/cfdgcn.py	@yangrongxinuser #1026 #1037
11	cylinder2d_unsteady.md	cylinder/2d_unsteady/cylinder2d_unsteady_Re100.py	@wufei2 #883
12	cylinder2d_unsteady_transformer_physx.md	cylinder/2d_unsteady/transformer_physx/train_transformer.py	@wufei2 #883
13	darcy2d.md	darcy/darcy2d.py	@wufei2 #900
14	deepcfd.md	deepcfd/deepcfd.py	@GoldenStain #994
15	ldc2d_steady.md	ldc/ldc2d_steady_Re10.py	@wufei2 #875
16	ldc2d_unsteady.md	ldc/ldc2d_unsteady_Re10.py	@wufei2 #875
17	labelfree_DNN_surrogate.md	pipe/poiseuille_flow.py	@jincheng23 @yangrongxinuser @PolaKuma #1030
18	nsfnet.md	aneurysm/aneurysm_flow.py
19	phycrnet.md	phycrnet/main.py	@Sylence8 #1029 @fangfangssj
20	shock_wave.md	shock_wave/shock_wave.py	@smallpoxscattered #890
21	tempoGAN.md	tempoGAN/tempoGAN.py	@smallpoxscattered #884
22	nsfnet4.md	nsfnet/VP_NSFNet4.py	@smallpoxscattered #864
23	viv.md	fsi/viv.py	@smallpoxscattered #832
24	biharmonic2d.md	biharmonic2d/biharmonic2d.py	@smallpoxscattered #858
25	bracket.md	bracket/bracket.py	@1want2sleep #878
26	control_arm.md	control_arm/forward_analysis.py	@smallpoxscattered #893
27	epnn.md	control_arm/inverse_parameter.py	@smallpoxscattered #893
28	phylstm.md	phylstm/phylstm2.pyphylstm/phylstm3.py	@jincheng23 @wwwuyan #1031
29	topopt.md	topopt/topopt.py	@NKNaN #793
30	heat_exchanger.md	heat_exchanger/heat_exchanger.py	@wufei2 #892
31	heat_pinn.md	heat_pinn/heat_pinn.py	@1want2sleep #926
32	phygeonet.md	phygeonet/heat_equation.py phygeonet/heat_equation_with_bc.py	@yangrongxinuser @ZHOU05030 @PolaKuma #1036
33	hpinns.md	hpinns/holography.py
34	fourcastnet.md	fourcastnet/train_pretrain.py fourcastnet/train_precip.py fourcastnet/train_finetune.py	@mayiru1021 #1129
35	nowcastnet.md	nowcastnet/nowcastnet.py	@smallpoxscattered #895 @kineast

认领方式

请大家以 comment 的形式认领任务，如：

【报名】：1、3、12-13

多个任务之间需要使用中文顿号分隔，报名多个连续任务可用横线表示，如 2-5
PR 提交格式：在 PR 的标题中以 【PPSCI Export&Infer No.】 开头，注明任务编号

看板信息

任务方向	任务数量	提交作品 / 任务认领	提交率	完成	完成率
快乐开源	35	33 / 33	94.29%	28	80.0%

统计信息

排名不分先后 @HydrogenSulfate (1) @GreatV (5) @wufei2 (7) @GoldenStain (1) @PolaKuma (2) @smallpoxscattered (8) @1want2sleep (2) @NKNaN (1) @mayiru1021 (1)

1. 背景

PaddleScience 套件为现有的 30+ 模型支持了一键训练、一键测试功能，但通过模型导出和python推理功能才能使得每个案例可以快速部署在各种设备上。

2. 收益

学习模型导出和python推理的基本流程，为案例代码添加模型导出和python推理功能

3. 开发流程

3.1 安装 PaddleScience

1. 安装 Paddle：https://paddlescience-docs.readthedocs.io/zh/latest/zh/install_setup/#13-paddlepaddle
2. 安装 PaddleScience(git 源码安装)：https://paddlescience-docs.readthedocs.io/zh/latest/zh/install_setup/#__tabbed_1_1

3.2 代码开发

以为 aneurysm 案例添加模型导出和 python 推理函数为例

def export(cfg: DictConfig):
    # set model
    model = ppsci.arch.MLP(**cfg.MODEL)

    # initialize solver
    solver = ppsci.solver.Solver(
        model,
        pretrained_model_path=cfg.INFER.pretrained_model_path,
    )
    # export model
    from paddle.static import InputSpec

    input_spec = [
        {key: InputSpec([None, 1], "float32", name=key) for key in model.input_keys},
    ]
    solver.export(input_spec, cfg.INFER.export_path)

1. 参考train/eval函数，以同样的方式实例化一个网络模型model，如果

2. 实例化一个solver，传入model、预训练模型路径变量cfg.INFER.pretrained_model_path（该路径设置为案例文档开头“模型评估命令”的EVAL.pretrained_model_path后的url值即可，如下所示）
   

3. 根据model的forwrad函数接受的输入格式，构造同样格式的input_spec。aneurysm 的输入格式为{"x": Tensor, "y": Tensor, "z": Tensor}，因此构造的input_spec为：{"x": InputSpec([None, 1], "float32", name="x"), "y": InputSpec([None, 1], "float32", name="y"), "z": InputSpec([None, 1], "float32", name="z")}

4. 调用solver.export导出模型至cfg.INFER.export_path路径下，导出成功后会打印：Inference model has been exported to: ./inference/aneurysm, including *.pdmodel, *.pdiparams and *.pdiparams.info files.
   至此模型导出函数export就完成了。接下来介绍如何撰写 python 推理代码inference函数

def inference(cfg: DictConfig):
    from deploy.python_infer import pinn_predictor

    predictor = pinn_predictor.PINNPredictor(cfg)
    eval_data_dict = reader.load_csv_file(
        cfg.EVAL_CSV_PATH,
        ("x", "y", "z", "u", "v", "w", "p"),
        {
            "x": "Points:0",
            "y": "Points:1",
            "z": "Points:2",
            "u": "U:0",
            "v": "U:1",
            "w": "U:2",
            "p": "p",
        },
    )
    input_dict = {
        "x": (eval_data_dict["x"] - cfg.CENTER[0]) * cfg.SCALE,
        "y": (eval_data_dict["y"] - cfg.CENTER[1]) * cfg.SCALE,
        "z": (eval_data_dict["z"] - cfg.CENTER[2]) * cfg.SCALE,
    }
    output_dict = predictor.predict(input_dict, cfg.INFER.batch_size)

    # mapping data to cfg.INFER.output_keys
    output_dict = {
        store_key: output_dict[infer_key]
        for store_key, infer_key in zip(cfg.MODEL.output_keys, output_dict.keys())
    }

    ppsci.visualize.save_vtu_from_dict(
        "./aneurysm_pred.vtu",
        {**input_dict, **output_dict},
        input_dict.keys(),
        cfg.MODEL.output_keys,
    )

1. 导入本模型对应的predictor，因为 aneurysm 的输入都是形状为[batch_size, channels]的Tensor，因此使用导入pinn_predictor即可；如果输入形状语义或者组织结构与之不同，则需要参照deploy/python_infer/pinn_predictor.py的PINNPredictor，写一个对应的XXXPredictor，再导入使用
2. 以cfg为参数，实例化predictor
3. 参考eval函数，构造输入数据input_dict
4. 调用predictor.predict接受输入数据并推理得到输出output_dict
5. 由于 python 推理的输出名字与原model的输出名字不同，因此需要将输出output_dict的键重新映射回正确的键上
6. 参照eval函数，用保存/打印等方式记录推理输出output_dict
   运行结果如下所示
   

如果在模型导出/推理时需要额外的超参数，则可以将这些超参数添加到案例对应 YAML 文件的 INFER 字段下，参考：https://github.com/PaddlePaddle/PaddleScience/pull/786/files#diff-1c33a8b95e9d0c72395763d5c906936652654533714e23211e2e2bb204d379f3R50-R55

3.3 文档添加 export&inference 运行命令

按照 3. 编写文档教程，安装文档渲染插件并渲染文档，用网页打开文档预览链接。

以aneurysm.md文档为例，添加红框内两处文本即可

添加后文档渲染结果如下所示

3.4 整理代码并提交PR

参考 PaddleScience文档-贡献指南-整理代码并提交

4. 参考资料

• [Fea] Support python inference #773
• [Refine] Refine yinglong code #786

[NKNaN]

【报名】：29

[GreatV]

【报名】：3

[GreatV]

【报名】：4

[GreatV]

【报名】：5