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.

  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.

  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:

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:

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.