【Hackathon 9th No.105】基于Paddle实现CoNFiLD流场生成模型 by ADream-ki · Pull Request #1230 · PaddlePaddle/PaddleScience

ADream-ki · 2025-10-19T15:01:01Z

PR types

Others

PR changes

Others

Describe

添加CoNFILD案例

…in confild.py, enhancing code documentation and usability.

paddle-bot · 2025-10-19T15:01:08Z

Thanks for your contribution!

…andling and loss logging. Update _update_ema function to work with parameter dictionaries. Enhance log_loss_dict to control loss aggregation during validation. Modify GaussianDiffusion class to ensure loss calculation includes valid_mse and vb terms correctly.

…dability and consistency. Update variable names for clarity, enhance gradient handling, and streamline loss logging. Fix tensor type conversions and ensure proper handling of model parameters during training and evaluation.

fix

…irectory exists before saving model parameters. This change improves file management during training by organizing saved models into a specified output directory.

…for dynamic adjustment while defaulting to 15698. This enhances flexibility in model evaluation.

HydrogenSulfate · 2025-11-17T04:32:51Z

docs/zh/examples/confild.md

+git clone https://github.com/PaddlePaddle/PaddleScience.git
+cd PaddleScience/examples/confild
+python confild.py mode=train


Suggested change

git clone https://github.com/PaddlePaddle/PaddleScience.git

cd PaddleScience/examples/confild

python confild.py mode=train

python confild.py mode=train

已修改，请review

HydrogenSulfate · 2025-11-17T04:34:16Z

docs/zh/examples/confild.md

+```python
+class Normalizer_ts(object):
+    def __init__(self, params=[], method="-11", dim=None):
+        self.params = params
+        self.method = method
+        self.dim = dim
+
+    def fit_normalize(self, data):
+        assert type(data) == paddle.Tensor
+        if len(self.params) == 0:
+            if self.method == "-11" or self.method == "01":
+                if self.dim is None:
+                    self.params = paddle.max(x=data), paddle.min(x=data)
+                else:
+                    self.params = (
+                        paddle.max(keepdim=True, x=data, axis=self.dim),
+                        paddle.argmax(keepdim=True, x=data, axis=self.dim),
+                    )[0], (
+                        paddle.min(keepdim=True, x=data, axis=self.dim),
+                        paddle.argmin(keepdim=True, x=data, axis=self.dim),
+                    )[
+                        0
+                    ]
+            elif self.method == "ms":
+                if self.dim is None:
+                    self.params = paddle.mean(x=data, axis=self.dim), paddle.std(
+                        x=data, axis=self.dim
+                    )
+                else:
+                    self.params = paddle.mean(
+                        x=data, axis=self.dim, keepdim=True
+                    ), paddle.std(x=data, axis=self.dim, keepdim=True)
+            elif self.method == "none":
+                self.params = None
+        return self.fnormalize(data, self.params, self.method)
+
+    def normalize(self, new_data):
+        if not new_data.place == self.params[0].place:
+            self.params = self.params[0].to(new_data.place), self.params[1].to(
+                new_data.place
+            )
+        return self.fnormalize(new_data, self.params, self.method)
+
+    def denormalize(self, new_data_norm):
+        if not new_data_norm.place == self.params[0].place:
+            self.params = self.params[0].to(new_data_norm.place), self.params[1].to(
+                new_data_norm.place
+            )
+        return self.fdenormalize(new_data_norm, self.params, self.method)
+
+    def get_params(self):
+        if self.method == "ms":
+            print("returning mean and std")
+        elif self.method == "01":
+            print("returning max and min")
+        elif self.method == "-11":
+            print("returning max and min")
+        elif self.method == "none":
+            print("do nothing")
+        return self.params
+
+    @staticmethod
+    def fnormalize(data, params, method):
+        if method == "-11":
+            return (data - params[1].to(data.place)) / (
+                params[0].to(data.place) - params[1].to(data.place)
+            ) * 2 - 1
+        elif method == "01":
+            return (data - params[1].to(data.place)) / (
+                params[0].to(data.place) - params[1].to(data.place)
+            )
+        elif method == "ms":
+            return (data - params[0].to(data.place)) / params[1].to(data.place)
+        elif method == "none":
+            return data
+
+    @staticmethod
+    def fdenormalize(data_norm, params, method):
+        if method == "-11":
+            return (data_norm + 1) / 2 * (
+                params[0].to(data_norm.place) - params[1].to(data_norm.place)
+            ) + params[1].to(data_norm.place)
+        elif method == "01":
+            return data_norm * (
+                params[0].to(data_norm.place) - params[1].to(data_norm.place)
+            ) + params[1].to(data_norm.place)
+        elif method == "ms":
+            return data_norm * params[1].to(data_norm.place) + params[0].to(
+                data_norm.place
+            )
+        elif method == "none":
+            return data_norm
+```


文档内的代码片段可以使用引用语法，引用.py文件里的代码

已修改，请review

HydrogenSulfate · 2025-11-17T04:34:38Z

docs/zh/examples/confild.md

+class SIRENAutodecoder_film(paddle.nn.Layer):
+    """
+    siren network with author decoding
+
+    Args:
+        input_keys (Tuple[str,...], optional): Key to get the input tensor from the dict.
+        output_keys (Tuple[str,...], optional): Key to save the output tensor into the dict.
+        in_coord_features (int, optional): Number of input coordinates features
+        in_latent_features (int, optional): Number of input latent features
+        out_features (int, optional): Number of output features
+        num_hidden_layers (int, optional): Number of hidden layers
+        hidden_features (int, optional): Number of hidden features
+        outermost_linear (bool, optional): Whether to use linear layer at the end. Defaults to False.
+        nonlinearity (str, optional): Nonlinearity to use. Defaults to "sine".
+        weight_init (Callable, optional): Weight initialization function. Defaults to None.
+        bias_init (Callable, optional): Bias initialization function. Defaults to None.
+        premap_mode (str, optional): Feature mapping mode. Defaults to None.
+
+    Examples:
+        >>> model = ppsci.arch.SIRENAutodecoder_film(
+                input_keys=["input1", "input2"],
+                output_keys=("output",),
+                in_coord_features=2,
+                in_latent_features=128,
+                out_features=3,
+                num_hidden_layers=10,
+                hidden_features=128,
+            )
+        >>> input_data = {"input1": paddle.randn([10, 2]), "input2": paddle.randn([10, 128])}
+        >>> out_dict = model(input_data)
+        >>> for k, v in out_dict.items():
+        ...     print(k, v.shape)
+        output [22, 918, 3]
+    """
+
+    def __init__(
+        self,
+        input_keys,
+        output_keys,
+        in_coord_features,
+        in_latent_features,
+        out_features,
+        num_hidden_layers,
+        hidden_features,
+        outermost_linear=False,
+        nonlinearity="sine",
+        weight_init=None,
+        bias_init=None,
+        premap_mode=None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.input_keys = input_keys
+        self.output_keys = output_keys
+
+        self.premap_mode = premap_mode
+        if self.premap_mode is not None:
+            self.premap_layer = FeatureMapping(
+                in_coord_features, mode=premap_mode, **kwargs
+            )
+            in_coord_features = self.premap_layer.dim
+        self.first_layer_init = None
+        self.nl, nl_weight_init, first_layer_init = NLS_AND_INITS[nonlinearity]
+        if weight_init is not None:
+            self.weight_init = weight_init
+        else:
+            self.weight_init = nl_weight_init
+        self.net1 = paddle.nn.LayerList(
+            sublayers=[BatchLinear(in_coord_features, hidden_features)]
+            + [
+                BatchLinear(hidden_features, hidden_features)
+                for i in range(num_hidden_layers)
+            ]
+            + [BatchLinear(hidden_features, out_features)]
+        )
+        self.net2 = paddle.nn.LayerList(
+            sublayers=[
+                BatchLinear(in_latent_features, hidden_features, bias_attr=False)
+                for i in range(num_hidden_layers + 1)
+            ]
+        )
+        if self.weight_init is not None:
+            self.net1.apply(self.weight_init)
+            self.net2.apply(self.weight_init)
+        if first_layer_init is not None:
+            self.net1[0].apply(first_layer_init)
+            self.net2[0].apply(first_layer_init)
+        if bias_init is not None:
+            self.net2.apply(bias_init)
+
+    def forward(self, input_data):
+        coords = input_data[self.input_keys[0]]
+        latents = input_data[self.input_keys[1]]
+        if self.premap_mode is not None:
+            x = self.premap_layer(coords)
+        else:
+            x = coords
+
+        for i in range(len(self.net1) - 1):
+            x = self.net1[i](x) + self.net2[i](latents)
+            x = self.nl(x)
+        x = self.net1[-1](x)
+        return {self.output_keys[0]: x}
+
+    def disable_gradient(self):
+        for param in self.parameters():
+            param.stop_gradient = not False


同上，引用

已修改，请review

HydrogenSulfate · 2025-11-17T04:34:41Z

docs/zh/examples/confild.md

+4.3 模型训练、评估
+完成上述设置之后，只需要将上述实例化的对象按照文档进行组合，然后启动训练、评估。
+```python
+def signal_train(cfg, normed_coords, normed_fois, spatio_axis, out_normalizer):


同上，引用

已修改，请review

HydrogenSulfate · 2025-11-17T04:34:49Z

docs/zh/examples/confild.md

+```
+
+## 5. 实验结果
+


实验结果没有？

忘了加上去了，请review

HydrogenSulfate · 2025-11-17T04:37:58Z

examples/confild/confild.py

+    # 显示图形
+    plt.show()


.show可以删除，图片保存即可

Suggested change

# 显示图形

plt.show()

已修改，请review

HydrogenSulfate · 2025-11-17T04:45:43Z

ppsci/arch/confild.py

commit之前请安装pre-commit，并且逐个对上传的文件调用pre-commit run进行格式化

HydrogenSulfate · 2025-11-17T04:46:21Z

ppsci/arch/__init__.py

    "LatentNO",
    "LatentNO_time",
    "LNO",
+    "LossType",


XXType不需要作为公开接口，可以删除

但是他在外部也是需要使用的，还是说让其作为一个变量传进去？

HydrogenSulfate · 2025-11-17T04:48:41Z

ppsci/arch/confild.py

+class BatchLinear(paddle.nn.Linear):
+    """
+    Batch-wise linear transformation layer that supports manual parameter injection.
+
+    This layer extends paddle.nn.Linear to allow passing parameters explicitly,
+    which is useful for meta-learning and hypernetwork applications.
+
+    Args:
+        in_features (int): Size of input features.
+        out_features (int): Size of output features.
+
+    Note:
+        - Weight shape: (out_features, in_features)
+        - Bias shape: (out_features,)
+    """
+
+    __doc__ = paddle.nn.Linear.__doc__
+
+    def forward(self, input, params=None):
+        """
+        Forward pass with optional external parameters.
+
+        Args:
+            input (paddle.Tensor): Input tensor of shape (..., in_features).
+            params (OrderedDict, optional): External parameters dict containing 'weight' and optionally 'bias'.
+                                           If None, uses internal parameters. Defaults to None.
+
+        Returns:
+            paddle.Tensor: Output tensor of shape (..., out_features).
+        """
+        if params is None:
+            params = OrderedDict(self.named_parameters())
+        bias = params.get("bias", None)
+        weight = params["weight"]
+
+        output = paddle.matmul(x=input, y=weight)
+        if bias is not None:
+            output += bias.unsqueeze(axis=-2)
+        return output
+


BatchLinear跟nn.Linear的区别是什么，如果没区别是不是可以删掉

前者在前向传播时可以动态替换参数

HydrogenSulfate · 2025-11-17T04:50:38Z

ppsci/arch/confild.py

+            self.centers = paddle.base.framework.EagerParamBase.from_tensor(
+                tensor=paddle.empty(
+                    shape=(rbf_out_features, in_features), dtype="float32"
+                )
+            )


较新的paddle可以直接用torch类似的语法：self.centers = nn.Paramter(...)

已经修改了，请review

ADream-ki · 2025-12-18T14:40:20Z

@HydrogenSulfate 我已经补全了MD文档了，请review

ADream-ki added 6 commits September 21, 2025 18:34

【Hackathon 9th No.105】CoNFiLD 论文复现

b19d520

Add UNet model implementation and Gaussian diffusion methods

ed154e5

update unet model

119f02c

delete duplicate key

411ca7c

fix

645cb4c

Add detailed docstrings for activation functions, layers, and models …

9bf9f56

…in confild.py, enhancing code documentation and usability.

paddle-bot bot added the contributor label Oct 19, 2025

luotao1 mentioned this pull request Oct 21, 2025

【Hackathon 9th】开源贡献个人挑战赛 PaddlePaddle/Paddle#74773

Closed

ADream-ki and others added 7 commits November 2, 2025 11:00

fix

f34c982

Merge branch 'H9NO105' into add_err_func

c4c25e8

Merge pull request #1 from wangguan1995/add_err_func

32e7bf7

fix

Enhance model saving functionality in confild.py by ensuring output d…

931b8ee

…irectory exists before saving model parameters. This change improves file management during training by organizing saved models into a specified output directory.

Make t_std configurable in evaluate function of confild.py, allowing …

46a5260

…for dynamic adjustment while defaulting to 15698. This enhances flexibility in model evaluation.

HydrogenSulfate requested changes Nov 17, 2025

View reviewed changes

ADream-ki and others added 5 commits December 4, 2025 21:29

feat:update code

3360c00

update confild.md

77df8ae

feat：update

31cb90b

update

751fced

Merge branch 'develop' into H9NO105

883ead0

luotao1 added the PaddlePaddle Hackathon label Dec 5, 2025

luotao1 assigned luotao1, KaiCHEN-HT and HydrogenSulfate Dec 5, 2025

ADream-ki and others added 3 commits December 6, 2025 00:38

update code

37cecac

update code style

70f4740

Merge branch 'develop' into H9NO105

bf64f27

PaddlePaddle locked and limited conversation to collaborators Dec 16, 2025

PaddlePaddle unlocked this conversation Dec 16, 2025

ADream-ki added 2 commits December 16, 2025 22:32

update code

8a6ec85

update result

cd2b32b

		```

		## 5. 实验结果

Conversation

ADream-ki commented Oct 19, 2025

PR types

PR changes

Describe

Uh oh!

paddle-bot bot commented Oct 19, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ADream-ki commented Dec 18, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants