Class Column

python/dgl/frame.py:167–568

A column is a compact store of features of multiple nodes/edges. It batches all the feature tensors together along the first dimension as one dense tensor. The column can optionally have an index tensor I. In this case, the i^th feature is stored in ``storage[index[i]]``. The c

class Column(TensorStorage):
    """A column is a compact store of features of multiple nodes/edges.

    It batches all the feature tensors together along the first dimension
    as one dense tensor.

    The column can optionally have an index tensor I.
    In this case, the i^th feature is stored in ``storage[index[i]]``.
    The column class implements a Copy-On-Read semantics -- the index
    select operation happens upon the first read of the feature data.
    This is useful when one extracts a subset of the feature data
    but wishes the actual index select happens on-demand.

    Parameters
    ----------
    storage : Tensor
        The feature data storage.
    scheme : Scheme, optional
        The scheme of the column. Will be inferred if not provided.
    index : Tensor, optional
        The row index to the feature data storage. None means an
        identity mapping.

    Attributes
    ----------
    storage : Tensor
        The storage tensor. The storage tensor may not be the actual data
        tensor of this column when the index tensor is not None.
        This typically happens when the column is extracted from another
        column using the `subcolumn` method.

        It can also be None, which may only happen when transmitting a
        not-yet-materialized subcolumn from a subprocess to the main process.
        In this case, the main process should already maintain the content of
        the storage, and is responsible for restoring the subcolumn's storage pointer.
    data : Tensor
        The actual data tensor of this column.
    scheme : Scheme
        The scheme of the column.
    index : Tensor
        Index tensor
    """

    def __init__(self, storage, *args, **kwargs):
        super().__init__(storage)
        self._init(*args, **kwargs)

    def __len__(self):
        """The number of features (number of rows) in this column."""
        if self.index is None:
            return F.shape(self.storage)[0]
        else:
            return len(self.index)

    @property
    def shape(self):
        """Return the scheme shape (feature shape) of this column."""
        return self.scheme.shape