【PPSCI Export&Infer No.24】 biharmonic2d (#858)

* [SCI Export&Infer No.24] biharmonic2d

* P[PSCI Export&Infer No.724] biharmonic2d fix

Author: smallpoxscattered

docs/zh/examples/biharmonic2d.md

@@ -14,6 +14,18 @@
     python biharmonic2d.py mode=eval EVAL.pretrained_model_path=https://paddle-org.bj.bcebos.com/paddlescience/models/biharmonic2d/biharmonic2d_pretrained.pdparams
     ```
 
+=== "模型导出命令"
+
+    ``` sh
+    python biharmonic2d.py mode=export
+    ```
+
+=== "模型推理命令"
+
+    ``` sh
+    python biharmonic2d.py mode=infer
+    ```
+
 | 预训练模型  | 指标 |
 |:--| :--|
 | [biharmonic2d_pretrained.pdparams](https://paddle-org.bj.bcebos.com/paddlescience/models/biharmonic2d/biharmonic2d_pretrained.pdparams) | l2_error: 0.02774 |

examples/biharmonic2d/biharmonic2d.py

@@ -39,7 +39,7 @@ def plotting(figname, output_dir, data, griddata_points, griddata_xi, boundary):
     for i, key in enumerate(data):
         plot_data = griddata(
             griddata_points,
-            data[key].numpy().flatten(),
+            data[key].flatten(),
             griddata_xi,
             method="cubic",
         )
@@ -343,7 +343,94 @@ def compute_outs(w, x, y):
     plotting(
         "eval_Mx_Mxy_My_Qx_Qy_w",
         cfg.output_dir,
-        outs,
+        {k: v.numpy() for k, v in outs.items()},
+        griddata_points,
+        griddata_xi,
+        boundary,
+    )
+
+
+def export(cfg: DictConfig):
+    from paddle import nn
+    from paddle.static import InputSpec
+
+    # set models
+    disp_net = ppsci.arch.MLP(**cfg.MODEL)
+
+    # load pretrained model
+    solver = ppsci.solver.Solver(
+        model=disp_net, pretrained_model_path=cfg.INFER.pretrained_model_path
+    )
+
+    class Wrapped_Model(nn.Layer):
+        def __init__(self, model):
+            super().__init__()
+            self.model = model
+
+        def forward(self, x):
+            model_out = self.model(x)
+            outs = self.compute_outs(model_out["u"], x["x"], x["y"])
+            return outs
+
+        def compute_outs(self, w, x, y):
+            D = cfg.E * (cfg.HEIGHT**3) / (12.0 * (1.0 - cfg.NU**2))
+            w_x2 = hessian(w, x)
+            w_y2 = hessian(w, y)
+            w_x_y = jacobian(jacobian(w, x), y)
+            M_x = -(w_x2 + cfg.NU * w_y2) * D
+            M_y = -(cfg.NU * w_x2 + w_y2) * D
+            M_xy = (1 - cfg.NU) * w_x_y * D
+            Q_x = -jacobian((w_x2 + w_y2), x) * D
+            Q_y = -jacobian((w_x2 + w_y2), y) * D
+            return {"Mx": M_x, "Mxy": M_xy, "My": M_y, "Qx": Q_x, "Qy": Q_y, "w": w}
+
+    solver.model = Wrapped_Model(solver.model)
+
+    # export models
+    input_spec = [
+        {key: InputSpec([None, 1], "float32", name=key) for key in disp_net.input_keys},
+    ]
+    solver.export(input_spec, cfg.INFER.export_path)
+
+
+def inference(cfg: DictConfig):
+    from deploy.python_infer import pinn_predictor
+
+    # set model predictor
+    predictor = pinn_predictor.PINNPredictor(cfg)
+
+    # generate samples
+    num_x = 201
+    num_y = 301
+    x_grad, y_grad = np.meshgrid(
+        np.linspace(
+            start=0, stop=cfg.LENGTH, num=num_x, endpoint=True, dtype=np.float32
+        ),
+        np.linspace(
+            start=0, stop=cfg.WIDTH, num=num_y, endpoint=True, dtype=np.float32
+        ),
+    )
+    x_faltten = x_grad.reshape(-1, 1)
+    y_faltten = y_grad.reshape(-1, 1)
+
+    output_dict = predictor.predict(
+        {"x": x_faltten, "y": y_faltten}, 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.INFER.output_keys, output_dict.keys())
+    }
+
+    # plotting
+    griddata_points = np.concatenate([x_faltten, y_faltten], axis=-1)
+    griddata_xi = (x_grad, y_grad)
+    boundary = [0, cfg.LENGTH, 0, cfg.WIDTH]
+    plotting(
+        "eval_Mx_Mxy_My_Qx_Qy_w",
+        cfg.output_dir,
+        output_dict,
         griddata_points,
         griddata_xi,
         boundary,
@@ -356,8 +443,14 @@ def main(cfg: DictConfig):
         train(cfg)
     elif cfg.mode == "eval":
         evaluate(cfg)
+    elif cfg.mode == "export":
+        export(cfg)
+    elif cfg.mode == "infer":
+        inference(cfg)
     else:
-        raise ValueError(f"cfg.mode should in ['train', 'eval'], but got '{cfg.mode}'")
+        raise ValueError(
+            f"cfg.mode should in ['train', 'eval', 'export', 'infer'], but got '{cfg.mode}'"
+        )
 
 
 if __name__ == "__main__":

examples/biharmonic2d/conf/biharmonic2d.yaml

@@ -11,6 +11,8 @@ hydra:
           - TRAIN.checkpoint_path
           - TRAIN.pretrained_model_path
           - EVAL.pretrained_model_path
+          - INFER.pretrained_model_path
+          - INFER.export_path
           - mode
           - output_dir
           - log_freq
@@ -72,3 +74,21 @@ EVAL:
   eval_with_no_grad: true
   batch_size:
     sup_validator: 128
+
+INFER:
+  pretrained_model_path: https://paddle-org.bj.bcebos.com/paddlescience/models/biharmonic2d/biharmonic2d_pretrained.pdparams
+  export_path: ./inference/biharmonic2d
+  pdmodel_path: ${INFER.export_path}.pdmodel
+  pdpiparams_path: ${INFER.export_path}.pdiparams
+  output_keys: ["Mx", "Mxy", "My", "Qx", "Qy", "w"]
+  device: gpu
+  engine: native
+  precision: fp32
+  onnx_path: ${INFER.export_path}.onnx
+  ir_optim: true
+  min_subgraph_size: 10
+  gpu_mem: 4000
+  gpu_id: 0
+  max_batch_size: 128
+  num_cpu_threads: 4
+  batch_size: 128