【Hackathon 8th No.11】DrivAerNet++ 论文复现 #1062

LilaKen · 2025-01-12T15:34:18Z

PR types

Others

PR changes

APIs、Docs

Describe

本PR的目标是复现RegPointNet网络，将DrivAerNet++处理成ppsci版API，并且加入对应的文档说明。

论文信息

年份	会议	作者	引用数	论文PDF
2024	Conference and Workshop on Neural Information Processing Systems (NeurIPS)	Mohamed Elrefaie, Florin Morar, Angela Dai, Faez Ahmed	4	DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks

现结果

预训练模型	神经网络	指标
DragPrediction_DrivAerNet_PointNet_r2_batchsize16_200epochs_100kpoints_tsne_NeurIPS_best_model.pdparams	RegPointNet	$R^2: 92%$

paddle-bot · 2025-01-12T15:34:23Z

Thanks for your contribution!

LilaKen · 2025-01-12T15:42:30Z

@HydrogenSulfate @wangguan1995 麻烦两位飞桨导师检查下该PR。

HydrogenSulfate · 2025-01-12T15:54:27Z

docs/zh/examples/drivaernetplusplus.md

+
+DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks(DrivAerNet++：一个包含计算流体力学模拟和深度学习基准的大规模多模态汽车数据集)
+
+## 论文信息:


这个末尾的冒号可以删掉，另一个文档也是

HydrogenSulfate · 2025-01-12T16:01:23Z

docs/zh/examples/drivaernetplusplus.md

+
+DrivAerNet++ 提供了一个用于汽车空气动力学性能预测的多模态数据集，任务目标是基于输入数据预测汽车的空气阻力系数（$C_d$）。
+
+##### 输入 ：


标题末尾的冒号请删除掉

HydrogenSulfate · 2025-01-12T16:03:12Z

examples/drivaernetplusplus/drivaernetplusplus.py

+
+
+def train(cfg: DictConfig):
+


多余的空行删除

HydrogenSulfate · 2025-01-12T16:03:52Z

examples/drivaernetplusplus/drivaernetplusplus.py

+            "input_keys": ("vertices",),
+            "label_keys": ("cd_value",),
+            "weight_keys": ("weight_keys",),


请实用cfg的配置

不太懂这里的意思

不太懂这里的意思

哦字打错了，是“使用”，就是这里不要用hard code的字符串，而是用cfg的属性

HydrogenSulfate · 2025-01-12T16:04:04Z

examples/drivaernetplusplus/drivaernetplusplus.py

+            "input_keys": ("vertices",),
+            "label_keys": ("cd_value",),
+            "weight_keys": ("weight_keys",),


HydrogenSulfate · 2025-01-12T16:06:51Z

ppsci/arch/regpointnet.py

+    Initialize the RegPointNet model for regression tasks with enhanced complexity,
+    including additional layers and residual connections.
+
+    Parameters:


docstring和现有代码保持一致，用Args而不是Parameters

__init__的docstring移动到类下方

HydrogenSulfate · 2025-01-12T16:07:22Z

ppsci/arch/regpointnet.py

+    """
+    Forward pass of the network.
+
+    Parameters:


HydrogenSulfate · 2025-01-12T16:07:37Z

ppsci/arch/regpointnet.py

+        x (Tensor): Input tensor of shape (batch_size, 3, num_points).
+
+    Returns:
+        Tensor: Output tensor of the predicted scalar value.


返回值类型请写完整，paddle.Tensor

HydrogenSulfate · 2025-01-12T16:08:13Z

ppsci/data/dataset/drivaernetplusplus_dataset.py

+        normalized_data = (data - mean_vals) / std_vals
+        return normalized_data
+
+    def mean_normalize(self, data: paddle.Tensor) -> paddle.Tensor:


这个函数有用到吗？

HydrogenSulfate · 2025-01-12T16:08:23Z

ppsci/data/dataset/drivaernetplusplus_dataset.py

+        normalized_data = (data - min_vals) / (max_vals - min_vals)
+        return normalized_data
+
+    def z_score_normalize(self, data: paddle.Tensor) -> paddle.Tensor:


这个函数有用到吗？

HydrogenSulfate · 2025-01-14T03:34:07Z

@LilaKen 有时间可以参考review建议，修改一下

HydrogenSulfate · 2025-01-15T06:43:40Z

@LilaKen commit时还麻烦再解决一下冲突

HydrogenSulfate · 2025-01-15T06:45:34Z

另外还有几个评论可能被缩进了，麻烦也看一下，

LilaKen · 2025-01-15T15:10:47Z

@LilaKen commit时还麻烦再解决一下冲突

已解决

LilaKen · 2025-01-15T15:10:58Z

另外还有几个评论可能被缩进了，麻烦也看一下，

已解决

HydrogenSulfate

另一个DrivAerNet文档也麻烦修改一下：

删除标题结尾冒号
请处理重复的的3.2章节

HydrogenSulfate · 2025-01-15T15:15:38Z

ppsci/arch/regpointnet.py

+        Returns:
+            paddle.Tensor: Output tensor of the predicted scalar value.


返回值类型不对

forward缺少type hint

HydrogenSulfate · 2025-01-15T15:16:00Z

ppsci/arch/regpointnet.py

+        Forward pass of the network.
+
+        Args:
+            x (Tensor): Input tensor of shape (batch_size, 3, num_points).


paddle.Tensor

HydrogenSulfate · 2025-01-15T15:16:23Z

ppsci/data/dataset/drivaernetplusplus_dataset.py

+                    f"Error loading point cloud from {load_path}: {e}"
+                ) from e
+
+    def __getitem__(self, idx: int, apply_augmentations: bool = True):


apply_augmentations这个参数没用到

HydrogenSulfate · 2025-01-15T15:16:30Z

ppsci/data/dataset/drivaernetplusplus_dataset.py

+            apply_augmentations (bool, optional): Whether to apply data augmentations. Defaults to True.
+
+        Returns:
+            Tuple[paddle.Tensor, paddle.Tensor]: The sample (point cloud) and its label (Cd value).


返回值类型不对

修改为Tuple[Dict[str, np.ndarray], Dict[str, np.ndarray], Dict[str, np.ndarray]]:

HydrogenSulfate · 2025-01-15T15:23:45Z

docs/zh/examples/drivaernetplusplus.md

+2. **批次大小的影响**：
+   - 如果使用不同的批次大小（batch size），R² 的结果可能会显著不同。例如，较小的批次可能会导致更大的方差，从而影响 R² 的计算。因此，基于批次的 R² 计算结果依赖于批次大小的选择，缺乏稳定性。
+
+```python  


```后面加一个空格，和其他地方保持一致

python和python格式化后还是```python

HydrogenSulfate · 2025-01-15T15:24:13Z

docs/zh/examples/drivaernetplusplus.md

整个文档是否能用 vscode 的markdown插件格式化一下？

已格式化

docs/zh/examples/drivaernetplusplus.md

HydrogenSulfate · 2025-01-15T15:28:03Z

ppsci/data/dataset/drivaernetplusplus_dataset.py

+            self.subset_indices = self.data_frame[
+                self.data_frame["Design"].isin(subset_ids)
+            ].index.tolist()
+            self.data_frame = self.data_frame.loc[self.subset_indices].reset_index(
+                drop=True
+            )


同，除打开文件外的操作请不要放到try内

HydrogenSulfate · 2025-01-15T15:28:59Z

mkdocs.yml

@@ -76,6 +76,8 @@ nav:
          - tempoGAN: zh/examples/tempoGAN.md
          - NSFNet4: zh/examples/nsfnet4.md
          - ViV: zh/examples/viv.md
+          - DrivAerNet: zh/examples/drivaernet.md


DrivAerNet在69行已经有了

DrivAerNetPlusPlus可以移动到70行

HydrogenSulfate · 2025-01-16T12:45:18Z

@LilaKen 有时间麻烦再改一下

LilaKen · 2025-01-16T15:10:04Z

@LilaKen 有时间麻烦再改一下

好的，可能需要过几天，手头上导师有些事情，我处理完把这个处理好，顺便也把DrivAerNet的重新pr更新一遍。

examples/drivaernetplusplus/drivaernetplusplus.py

HydrogenSulfate · 2025-01-26T16:31:48Z

mkdocs.yml

本文件内无关内容包括缩进请不要改动，否则本地无法构建文档

HydrogenSulfate · 2025-01-26T16:38:56Z

ppsci/arch/regpointnet.py

+        # Ensure x is in the shape (bs, C, L)
+        if len(x.shape) != 3:
+            raise ValueError(
+                f"Input tensor should have shape (batch_size, C, L), but got {x.shape}"
+            )
+
+        # Initialize an empty list to store the processed samples
+        processed_samples = []
+
+        # Apply data augmentation and normalization for each sample in the batch
+        augmentation = DataAugmentation()
+        for i in range(bs):
+            sample = x[i].numpy()  # Convert to numpy array for data augmentation
+            sample = augmentation.translate_pointcloud(sample)
+            sample = augmentation.jitter_pointcloud(sample)
+            sample = self.min_max_normalize(sample)
+            processed_samples.append(sample)
+
+        # Stack the processed samples back into a batch tensor
+        x_processed = paddle.to_tensor(np.stack(processed_samples, axis=0))
+
+        # Ensure the processed tensor has the same shape as the original input
+        if x_processed.shape != x.shape:
+            raise ValueError(
+                f"Processed tensor shape {x_processed.shape} does not match original input shape {x.shape}"
+            )


为什么要把数据增强代码写到模型前向里？

不好意思，之前找复现bug的时候挪动了一下，已经重新调回数据集加载py文件中。

Co-authored-by: HydrogenSulfate <[email protected]>

HydrogenSulfate · 2025-01-30T17:53:59Z

docs/zh/examples/drivaernetplusplus.md

+```python
+    # 源代码DeepSurrogates/train_RegPointNet.py中的R²计算


这个代码块不需要整体缩进4个空格吧?

已取消缩进

HydrogenSulfate · 2025-01-30T17:54:42Z

docs/zh/examples/drivaernetplusplus.md

+4. 启动 Globus Connect 个人版。由于第一次运行，因此必须先完成设置，然后才能运行完整的应用程序。
+
+   ```sh
+   $ ./globusconnectpersonal


sh命令里的美元符号建议删掉，全文检查一下

HydrogenSulfate · 2025-01-30T17:56:10Z

docs/zh/examples/drivaernetplusplus.md

+## 2. 问题定义
+
+#### 2.1 数据集呈现


二级标题后应该跟着的是三级标题，三个井号，文档整体检查一下标题级数的正确性

HydrogenSulfate · 2025-01-30T17:56:57Z

docs/zh/examples/drivaernetplusplus.md

+
+为了预测汽车的空气阻力系数 ($C_d$)，本研究采用基于深度学习的回归方法，使用两种点云处理模型 **RegDGCNN** 和 **RegPointNet**，分别从输入数据中提取几何特征并完成回归任务。这些模型能够高效处理 3D 点云数据，并结合设计参数，构建端到端的预测框架。
+
+##### **1. RegDGCNN**


Suggested change

##### **1. RegDGCNN**

1. RegDGCNN

docs/zh/examples/drivaernetplusplus.md

Co-authored-by: HydrogenSulfate <[email protected]>

LilaKen and others added 3 commits January 10, 2025 15:57

Update ReduceOnPlateau lr_scheduler.md

8c3438a

Merge branch 'PaddlePaddle:develop' into develop

4ca8511

support drivaernetplusplus

a3768f7

paddle-bot bot added the contributor label Jan 12, 2025

amend older error arch.md

2a6d9dd

luotao1 mentioned this pull request Jan 13, 2025

【Hackathon 7th】开源贡献个人挑战赛（偷跑版） PaddlePaddle/Paddle#67603

Closed

HydrogenSulfate reviewed Jan 13, 2025

View reviewed changes

luotao1 added the PaddlePaddle Hackathon label Jan 13, 2025

luotao1 assigned luotao1 and wangguan1995 Jan 13, 2025

luotao1 changed the title ~~【Hackathon 7th PPSCI No.3】NO.3 DrivAerNet++ 论文复现~~ 【Hackathon 8th No.11】DrivAerNet++ 论文复现 Jan 14, 2025

luotao1 mentioned this pull request Jan 14, 2025

【Hackathon 8th】开源贡献个人挑战赛（尝鲜版） PaddlePaddle/Paddle#70746

Open

LilaKen added 2 commits January 15, 2025 00:18

amend some question

b9caca5

update regpintnet.py

2d2d13f

amend regpointnet and drivaernetplusplus.py

e798771

HydrogenSulfate reviewed Jan 15, 2025

View reviewed changes

LilaKen added 5 commits January 26, 2025 02:04

update code and format markdown

7c98038

update code and format markdown

3b872d2

update code and format markdown

0e0acc0

update md

657b0ab

update code and format markdown

79facfa

LilaKen force-pushed the DrivAerNet++ branch from 0ff627a to 79facfa Compare January 25, 2025 18:33

LilaKen added 2 commits January 26, 2025 02:36

Merge branch 'develop' into DrivAerNet++

7125eaa

Update drivaernetplusplus_dataset.py

4714ac8

HydrogenSulfate reviewed Jan 26, 2025

View reviewed changes

LilaKen and others added 6 commits January 28, 2025 01:07

Update examples/drivaernetplusplus/drivaernetplusplus.py

beb8e38

Co-authored-by: HydrogenSulfate <[email protected]>

Update examples/drivaernetplusplus/drivaernetplusplus.py

bb45ad0

Co-authored-by: HydrogenSulfate <[email protected]>

Update examples/drivaernetplusplus/drivaernetplusplus.py

917ea4e

Co-authored-by: HydrogenSulfate <[email protected]>

move data_augmentation to load dataset file

9c0cc68

sysc the apply_augmentation in markdown file

a47e529

sysc the apply_augmentation in markdown file

a3f6a67

LilaKen force-pushed the DrivAerNet++ branch from f33670f to a3f6a67 Compare January 27, 2025 18:06

pre-commit file drivaernetplusplus.py

18c0be5

HydrogenSulfate reviewed Feb 2, 2025

View reviewed changes

LilaKen and others added 3 commits February 10, 2025 19:22

Update docs/zh/examples/drivaernetplusplus.md

33c73bd

Co-authored-by: HydrogenSulfate <[email protected]>

Merge branch 'develop' into DrivAerNet++

7a2a852

update drivaernetplusplus.md

40c61d9


		DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks(DrivAerNet++：一个包含计算流体力学模拟和深度学习基准的大规模多模态汽车数据集)

		## 论文信息:


		DrivAerNet++ 提供了一个用于汽车空气动力学性能预测的多模态数据集，任务目标是基于输入数据预测汽车的空气阻力系数（$C_d$）。

		##### 输入：

		Returns:
		paddle.Tensor: Output tensor of the predicted scalar value.

		```python
		# 源代码DeepSurrogates/train_RegPointNet.py中的R²计算


		为了预测汽车的空气阻力系数 ($C_d$)，本研究采用基于深度学习的回归方法，使用两种点云处理模型 RegDGCNN 和 RegPointNet，分别从输入数据中提取几何特征并完成回归任务。这些模型能够高效处理 3D 点云数据，并结合设计参数，构建端到端的预测框架。

		##### 1. RegDGCNN

【Hackathon 8th No.11】DrivAerNet++ 论文复现 #1062

Are you sure you want to change the base?

【Hackathon 8th No.11】DrivAerNet++ 论文复现 #1062

Conversation

LilaKen commented Jan 12, 2025

PR types

PR changes

Describe

论文信息

现结果

paddle-bot bot commented Jan 12, 2025

LilaKen commented Jan 12, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

HydrogenSulfate commented Jan 14, 2025

HydrogenSulfate commented Jan 15, 2025

HydrogenSulfate commented Jan 15, 2025

LilaKen commented Jan 15, 2025

LilaKen commented Jan 15, 2025

HydrogenSulfate left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

HydrogenSulfate commented Jan 16, 2025

LilaKen commented Jan 16, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment