图数据集的加载

原文参考官方文档:

https://pytorch-geometric.readthedocs.io/en/latest/modules/loader.html

torch_geometric.loader 库中, 该库中包含了多种 图数据集的 加载方式,

这里主要介绍 DenseDataLoader and DataLoader 这两者之间的区别;

Difference between DenseDataLoader and DataLoader in PyTorch Geometric

1. DenseDataLoader:

  • Purpose: Specifically designed for loading batches of dense graph data where all graph attributes have the same shape.

  • Stacking: Stacks all graph attributes in a new dimension, which means that all graph data needs to be dense (i.e., have the same shape).

  • Use Case: Ideal for situations where the adjacency matrices and feature matrices of graphs in the dataset are of consistent size and can be stacked without any padding or truncation.

  • Implementation : Uses a custom collate_fn that stacks all attributes of the graph data objects into a new dimension. This is suitable for dense graph data.

    python 复制代码
    class DenseDataLoader(torch.utils.data.DataLoader):
        ...
        def __init__(self, dataset: Union[Dataset, List[Data]],
                     batch_size: int = 1, shuffle: bool = False, **kwargs):
            kwargs.pop('collate_fn', None)  # Ensure no other collate function is used
            super().__init__(dataset, batch_size=batch_size, shuffle=shuffle,
                             collate_fn=collate_fn, **kwargs)

2. DataLoader:

  • Purpose: General-purpose data loader for PyTorch Geometric datasets. It can handle both homogeneous and heterogeneous graph data.

  • Flexibility: Can handle varying graph sizes and structures by merging data objects into mini-batches. Suitable for heterogeneous data where graph attributes may differ in shape and size.

  • Collate Function : Uses a custom collate function, Collater, which can handle different types of data elements (e.g., BaseData, Tensor, etc.). This function is versatile and can manage the complexity of heterogeneous graph data.

  • Use Case: Ideal for most graph data scenarios, especially when graphs vary in size and shape, and when working with both homogeneous and heterogeneous data.

    python 复制代码
    class DataLoader(torch.utils.data.DataLoader):
        ...
        def __init__(
            self,
            dataset: Union[Dataset, Sequence[BaseData], DatasetAdapter],
            batch_size: int = 1,
            shuffle: bool = False,
            follow_batch: Optional[List[str]] = None,
            exclude_keys: Optional[List[str]] = None,
            **kwargs,
        ):
            kwargs.pop('collate_fn', None)  # Ensure no other collate function is used
            super().__init__(
                dataset,
                batch_size,
                shuffle,
                collate_fn=Collater(dataset, follow_batch, exclude_keys),
                **kwargs,
            )

Key Differences

  1. Data Shape Consistency:

    • DenseDataLoader: Requires all graph attributes to have the same shape.
    • DataLoader: Can handle variable graph sizes and shapes.
  2. Batching Mechanism:

    • DenseDataLoader: Stacks all attributes into a new dimension, suitable for dense data.
    • DataLoader : Uses the Collater class to handle complex data batching, suitable for heterogeneous and variable-sized graph data.
  3. Use Cases:

    • DenseDataLoader: Best for datasets with consistent graph sizes and shapes.
    • DataLoader: Best for general-purpose graph data loading, especially with varying graph structures.

Practical Example of Each Loader

DenseDataLoader Example:

python 复制代码
from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DenseDataLoader

# Load a dataset where all graphs have the same number of nodes
dataset = TUDataset(root='/tmp/ENZYMES', name='ENZYMES')

# Create a DenseDataLoader
loader = DenseDataLoader(dataset, batch_size=32, shuffle=True)

for batch in loader:
    print(batch)
    break

DataLoader Example:

python 复制代码
from torch_geometric.datasets import TUDataset
from torch_geometric.loader import DataLoader

# Load a dataset with graphs of varying sizes
dataset = TUDataset(root='/tmp/ENZYMES', name='ENZYMES')

# Create a DataLoader
loader = DataLoader(dataset, batch_size=32, shuffle=True)

for batch in loader:
    print(batch)
    break

Summary

  • Use DenseDataLoader when working with datasets where all graphs have the same size and shape.
  • Use DataLoader for more flexible and general-purpose graph data loading, especially when dealing with variable graph structures.
相关推荐
泯泷43 分钟前
别再傻傻地喂文件了,大厂都在偷偷给AI建这种“专属图书馆”
人工智能
oak隔壁找我1 小时前
一款运行在 DSH(DeepSeek Harness)中的类 Postman HTTP 接口调试插件
人工智能·程序员
高洁011 小时前
数字孪生与计算机视觉:让机器看懂现场
人工智能·python·机器学习·transformer·迁移学习
Csvn2 小时前
如果明天供应商涨价 3 倍?三层解耦,换模型只改一行配置(C05)
人工智能·aigc·agent
在所不辞兄3 小时前
为什么数字孪生驱动的产线标定需要在线更新机制,而不能一次标定后长期沿用?
人工智能·神经网络·其他·算法·机器学习
锋行天下3 小时前
langgraph 全栈项目(三)
前端·人工智能·后端
黄啊码4 小时前
【黄啊码】最近面了十几个想转 FDE 的程序员,一个都没招
人工智能
锋行天下4 小时前
langgraph 全栈项目(二)
前端·人工智能·后端
大模型真好玩4 小时前
用 Seedance 2.5 Flash 实现1分钟AI短剧,满足孩子的超级英雄对战梦
人工智能·aigc·deepseek
FengY_HYY5 小时前
从人类组织到人类&大模型组织,我的位置在哪里?
人工智能·程序员·llm