[Doc] Add tutorial and fix docs

docs/zh/development.md

@@ -116,15 +116,15 @@ model = ppsci.arch.MLP(("x", "y"), ("u", "v", "p"), 9, 50, "tanh")
 
         ``` py
         --8<--
-        ppsci/arch/mlp.py:73:138
+        ppsci/arch/mlp.py:86:151
         --8<--
         ```
 
     === "MLP.forward"
 
         ``` py
         --8<--
-        ppsci/arch/mlp.py:140:167
+        ppsci/arch/mlp.py:153:180
         --8<--
         ```
 

docs/zh/examples/biharmonic2d.md

@@ -114,17 +114,17 @@ examples/biharmonic2d/biharmonic2d.py:93:95
 
 ``` py linenums="97"
 --8<--
-examples/biharmonic2d/biharmonic2d.py:97:108
+examples/biharmonic2d/biharmonic2d.py:97:106
 --8<--
 ```
 
 #### 3.4.1 内部约束
 
 以作用在背板内部点的 `InteriorConstraint` 为例，代码如下：
 
-``` py linenums="207"
+``` py linenums="205"
 --8<--
-examples/biharmonic2d/biharmonic2d.py:207:216
+examples/biharmonic2d/biharmonic2d.py:205:214
 --8<--
 ```
 
@@ -158,19 +158,19 @@ examples/biharmonic2d/conf/biharmonic2d.yaml:60:62
 
 #### 3.4.2 边界约束
 
-如 [2 问题定义](#2) 中所述，$x=0$ 处的挠度 $w$ 为 0，有如下边界条件，其他 7 个边界条件也与之类似：
+如 [2. 问题定义](#2) 中所述，$x=0$ 处的挠度 $w$ 为 0，有如下边界条件，其他 7 个边界条件也与之类似：
 
-``` py linenums="111"
+``` py linenums="108"
 --8<--
-examples/biharmonic2d/biharmonic2d.py:111:120
+examples/biharmonic2d/biharmonic2d.py:108:118
 --8<--
 ```
 
 在方程约束、边界约束构建完毕之后，以刚才的命名为关键字，封装到一个字典中，方便后续访问。
 
-``` py linenums="217"
+``` py linenums="215"
 --8<--
-examples/biharmonic2d/biharmonic2d.py:217:228
+examples/biharmonic2d/biharmonic2d.py:215:226
 --8<--
 ```
 
@@ -198,19 +198,19 @@ examples/biharmonic2d/conf/biharmonic2d.yaml:46:56
 
 完成上述设置之后，只需要将上述实例化的对象按顺序传递给 `ppsci.solver.Solver`，然后启动训练，注意两个优化过程需要分别构建 `Solver`。
 
-``` py linenums="230"
+``` py linenums="228"
 --8<--
-examples/biharmonic2d/biharmonic2d.py:230:269
+examples/biharmonic2d/biharmonic2d.py:228:267
 --8<--
 ```
 
 ### 3.8 模型评估和可视化
 
 训练完成后，可以在 `eval` 模式中对训练好的模型进行评估和可视化。由于案例的特殊性，不需构建评估器和可视化器，而是使用自定义代码。
 
-``` py linenums="272"
+``` py linenums="270"
 --8<--
-examples/biharmonic2d/biharmonic2d.py:272:352
+examples/biharmonic2d/biharmonic2d.py:270:350
 --8<--
 ```
 

docs/zh/tutorials.md

@@ -0,0 +1,6 @@
+# 学习资料
+
+## 教程课件
+
+- [深度学习技术与科学计算](https://aistudio.baidu.com/course/introduce/29929?sharedType=1&sharedUserId=438690&ts=1705731573142)
+- [飞桨AI for Science流体力学公开课第一期](https://aistudio.baidu.com/course/introduce/27926?sharedType=1&sharedUserId=438690&ts=1705892946215)

examples/biharmonic2d/biharmonic2d.py

@@ -103,8 +103,6 @@ def train(cfg: DictConfig):
             "drop_last": True,
             "shuffle": True,
         },
-        "num_workers": 0,
-        "auto_collation": False,
     }
 
     # set constraint

examples/biharmonic2d/conf/biharmonic2d.yaml

@@ -23,7 +23,7 @@ hydra:
 mode: train # running mode: train/eval
 seed: 2023
 output_dir: ${hydra:run.dir}
-log_freq: 200
+log_freq: 20
 
 # set working condition
 E: 201880.0e+6  # Pa = N/m2

examples/bubble/bubble.py

@@ -209,7 +209,7 @@ def transform_out(in_, out):
         NPOINT_PDE * NTIME_PDE, evenly=True
     )
 
-    pred_norm = solver.predict(visu_mat, no_grad=False, return_numpy=True)
+    pred_norm = solver.predict(visu_mat, None, 4096, no_grad=False, return_numpy=True)
     # inverse normalization
     p_pred = pred_norm["p"].reshape([NTIME_PDE, NPOINT_PDE]).T
     u_pred = pred_norm["u"].reshape([NTIME_PDE, NPOINT_PDE]).T
@@ -362,7 +362,9 @@ def transform_out(in_, out):
         NPOINT_PDE * NTIME_PDE, evenly=True
     )
 
-    pred_norm = solver.predict(visu_mat, None, 4096, no_grad=False, return_numpy=True)
+    pred_norm = solver.predict(
+        visu_mat, None, 4096 * 2, no_grad=False, return_numpy=True
+    )
     # inverse normalization
     p_pred = pred_norm["p"].reshape([NTIME_PDE, NPOINT_PDE]).T
     u_pred = pred_norm["u"].reshape([NTIME_PDE, NPOINT_PDE]).T

mkdocs.yml

@@ -34,6 +34,7 @@ nav:
       - 功能介绍: zh/overview.md
       - 安装使用: zh/install_setup.md
       - 快速开始: zh/quickstart.md
+      - 学习资料: zh/tutorials.md
   - 经典案例:
     - " ":
       - 数学(AI for Math):

ppsci/arch/activation.py

@@ -75,6 +75,12 @@ def forward(self, x):
 
 
 class Silu(nn.Layer):
+    """
+    FIXME: This activation function is a workaround for the potential occurrence of NaNs
+    during the computation of the native SiLU function via using x*sigmoid(x) instead of
+    silu(x)
+    """
+
     def __init__(self):
         super().__init__()
 

ppsci/arch/afno.py

@@ -529,9 +529,8 @@ def forward_tensor(self, x):
 
         return x
 
-    def split_to_dict(
-        self, data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]
-    ):
+    @staticmethod
+    def split_to_dict(data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]):
         return {key: data_tensors[i] for i, key in enumerate(keys)}
 
     def forward(self, x):
@@ -653,9 +652,8 @@ def forward_tensor(self, x):
         x = self.act(x)
         return x
 
-    def split_to_dict(
-        self, data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]
-    ):
+    @staticmethod
+    def split_to_dict(data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]):
         return {key: data_tensors[i] for i, key in enumerate(keys)}
 
     def forward(self, x):

ppsci/arch/base.py

@@ -60,8 +60,9 @@ def num_params(self) -> int:
                 logger.warning(f"{name} has no attribute 'shape'")
         return num
 
+    @staticmethod
     def concat_to_tensor(
-        self, data_dict: Dict[str, paddle.Tensor], keys: Tuple[str, ...], axis=-1
+        data_dict: Dict[str, paddle.Tensor], keys: Tuple[str, ...], axis=-1
     ) -> Tuple[paddle.Tensor, ...]:
         """Concatenate tensors from dict in the order of given keys.
 
@@ -95,8 +96,9 @@ def concat_to_tensor(
         data = [data_dict[key] for key in keys]
         return paddle.concat(data, axis)
 
+    @staticmethod
     def split_to_dict(
-        self, data_tensor: paddle.Tensor, keys: Tuple[str, ...], axis=-1
+        data_tensor: paddle.Tensor, keys: Tuple[str, ...], axis=-1
     ) -> Dict[str, paddle.Tensor]:
         """Split tensor and wrap into a dict by given keys.
 

ppsci/arch/embedding_koopman.py

@@ -177,9 +177,8 @@ def forward_tensor(self, x):
 
         return (pred_data[:, :-1, :], recover_data, k_matrix)
 
-    def split_to_dict(
-        self, data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]
-    ):
+    @staticmethod
+    def split_to_dict(data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]):
         return {key: data_tensors[i] for i, key in enumerate(keys)}
 
     def forward(self, x):
@@ -482,9 +481,8 @@ def forward_tensor(self, states, visc):
 
         return (pred_data[:, :-1], recover_data, k_matrix)
 
-    def split_to_dict(
-        self, data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]
-    ):
+    @staticmethod
+    def split_to_dict(data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]):
         return {key: data_tensors[i] for i, key in enumerate(keys)}
 
     def forward(self, x):

ppsci/arch/epnn.py

@@ -66,10 +66,13 @@ class Epnn(base.Arch):
     Examples:
         >>> import ppsci
         >>> ann_node_sizes_state = [1]
-        >>> model = ppsci.arch.Epnn(("x",), ("y",), node_sizes=node_sizes_state,
-                                    activations=("leaky_relu"),
-                                    drop_p=0.0
-                                    )
+        >>> model = ppsci.arch.Epnn(
+        ...     ("x",),
+        ...     ("y",),
+        ...     node_sizes=ann_node_sizes_state,
+        ...     activations=("leaky_relu"),
+        ...     drop_p=0.0,
+        ... )
     """
 
     def __init__(

ppsci/arch/gan.py

@@ -352,8 +352,9 @@ def forward(self, x):
 
         return y
 
+    @staticmethod
     def split_to_dict(
-        self, data_list: List[paddle.Tensor], keys: Tuple[str, ...]
+        data_list: List[paddle.Tensor], keys: Tuple[str, ...]
     ) -> Dict[str, paddle.Tensor]:
         """Overwrite of split_to_dict() method belongs to Class base.Arch.
 

ppsci/arch/nowcastnet.py

@@ -62,9 +62,8 @@ def __init__(
         sample_tensor = paddle.zeros(shape=[1, 1, self.image_height, self.image_width])
         self.grid = make_grid(sample_tensor)
 
-    def split_to_dict(
-        self, data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]
-    ):
+    @staticmethod
+    def split_to_dict(data_tensors: Tuple[paddle.Tensor, ...], keys: Tuple[str, ...]):
         return {key: data_tensors[i] for i, key in enumerate(keys)}
 
     def forward(self, x):

ppsci/arch/physx_transformer.py

@@ -367,7 +367,8 @@ def forward_eval(self, x):
         outputs = self.generate(input_embeds)
         return (outputs[:, 1:],)
 
-    def split_to_dict(self, data_tensors, keys):
+    @staticmethod
+    def split_to_dict(data_tensors, keys):
         return {key: data_tensors[i] for i, key in enumerate(keys)}
 
     def forward(self, x):

ppsci/data/__init__.py

@@ -128,7 +128,10 @@ def build_dataloader(_dataset, cfg):
             collate_fn=collate_fn,
         )
     else:
-        if cfg.get("auto_collation", True) is False:
+        if (
+            cfg.get("auto_collation", not getattr(_dataset, "batch_index", False))
+            is False
+        ):
             # 1. wrap batch_sampler again into BatchSampler for disabling auto collation,
             # which can speed up the process of batch samples indexing from dataset. See
             # details at: https://discuss.pytorch.org/t/efficiency-of-dataloader-and-collate-for-large-array-like-datasets/59569/8
@@ -140,10 +143,14 @@ def build_dataloader(_dataset, cfg):
                 )
             # 2. disable auto collation by given identity collate_fn which return the first
             # (also the only) batch data in batch list, or there will be a redundant
-            # axis at the first dimension returned by dataloader. It is step is necessary
+            # axis at the first dimension returned by dataloader. This step is necessary
             # because paddle do not support 'sampler' as instantiation argument of 'io.DataLoader'
             collate_fn = lambda batch: batch[0]  # noqa: E731
-            logger.info("Auto collation is disabled to speed up batch sampling")
+            _DEFAULT_NUM_WORKERS = 0
+            logger.info(
+                "Auto collation is disabled and set num_workers to "
+                f"{_DEFAULT_NUM_WORKERS} to speed up batch sampling."
+            )
 
         dataloader_ = io.DataLoader(
             dataset=_dataset,

ppsci/data/dataset/airfoil_dataset.py

@@ -109,6 +109,9 @@ class MeshAirfoilDataset(io.Dataset):
         ... )  # doctest: +SKIP
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = False
+
     use_pgl: bool = True
 
     def __init__(

ppsci/data/dataset/array_dataset.py

@@ -41,6 +41,9 @@ class NamedArrayDataset(io.Dataset):
         >>> dataset = ppsci.data.dataset.NamedArrayDataset(input, output, weight)
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = True
+
     def __init__(
         self,
         input: Dict[str, np.ndarray],
@@ -91,6 +94,9 @@ class IterableNamedArrayDataset(io.IterableDataset):
         >>> dataset = ppsci.data.dataset.IterableNamedArrayDataset(input, label, weight)
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = False
+
     def __init__(
         self,
         input: Dict[str, np.ndarray],

ppsci/data/dataset/csv_dataset.py

@@ -54,6 +54,9 @@ class CSVDataset(io.Dataset):
         ... )  # doctest: +SKIP
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = True
+
     def __init__(
         self,
         file_path: str,
@@ -176,6 +179,9 @@ class IterableCSVDataset(io.IterableDataset):
         ... )  # doctest: +SKIP
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = False
+
     def __init__(
         self,
         file_path: str,

ppsci/data/dataset/cylinder_dataset.py

@@ -55,6 +55,9 @@ class MeshCylinderDataset(io.Dataset):
         ... )  # doctest: +SKIP
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = False
+
     use_pgl: bool = True
 
     def __init__(

ppsci/data/dataset/era5_dataset.py

@@ -51,6 +51,9 @@ class ERA5Dataset(io.Dataset):
         ... )  # doctest: +SKIP
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = False
+
     def __init__(
         self,
         file_path: str,

ppsci/data/dataset/mat_dataset.py

@@ -54,6 +54,9 @@ class MatDataset(io.Dataset):
         ... )  # doctest: +SKIP
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = True
+
     def __init__(
         self,
         file_path: str,
@@ -176,6 +179,9 @@ class IterableMatDataset(io.IterableDataset):
         ... )  # doctest: +SKIP
     """
 
+    # Whether support batch indexing for speeding up fetching process.
+    batch_index: bool = False
+
     def __init__(
         self,
         file_path: str,