3、TENSOR VIEWS

布满荆棘的人生 2022-10-08 15:29 127阅读 0赞

PyTorch 允许张量是现有张量的视图。视图张量与其基础张量共享相同的基础数据。 支持 View 避免了显式数据复制，从而使我们能够进行快速且内存高效的整形、切片和元素操作。

例如，要获取现有张量 t 的视图，可以调用 t.view(...)。

>>> t = torch.rand(4, 4)
    >>> b = t.view(2, 8)
    >>> t.storage().data_ptr() == b.storage().data_ptr()  # `t` and `b` share the same underlying data.
    True
    # Modifying view tensor changes base tensor as well.
    >>> b[0][0] = 3.14
    >>> t[0][0]
    tensor(3.14)

由于视图与其基本张量共享基础数据，如果您在视图中编辑数据，它也会反映在基本张量中。

通常，PyTorch 操作会返回一个新的张量作为输出，例如 add（）。但是在视图操作的情况下，输出是输入张量的视图，以避免不必要的数据复制。创建视图时不会发生数据移动，视图张量只是改变了它解释相同数据的方式。考虑连续张量可能会产生不连续的张量。用户应特别注意，因为连续性可能会对性能产生隐含的影响。 transpose() 是一个常见的例子。

>>> base = torch.tensor([[0, 1],[2, 3]])
    >>> base.is_contiguous()
    True
    >>> t = base.transpose(0, 1)  # `t` is a view of `base`. No data movement happened here.
    # View tensors might be non-contiguous.
    >>> t.is_contiguous()
    False
    # To get a contiguous tensor, call `.contiguous()` to enforce
    # copying data when `t` is not contiguous.
    >>> c = t.contiguous()

作为参考，以下是 PyTorch 中视图操作的完整列表：

基本切片和索引操作，例如 tensor\[0, 2:, 1:7:2\] 返回基本张量的视图，请参见下面的注释。

*  [`as_strided()`][as_strided]
 *  [`detach()`][detach]
 *  [`diagonal()`][diagonal]
 *  [`expand()`][expand]
 *  [`expand_as()`][expand_as]
 *  [`movedim()`][movedim]
 *  [`narrow()`][narrow]
 *  [`permute()`][permute]
 *  [`select()`][select]
 *  [`squeeze()`][squeeze]
 *  [`transpose()`][transpose]
 *  [`t()`][t]
 *  [`T`][T]
 *  [`real`][real]
 *  [`imag`][imag]
 *  `view_as_real()`
 *  `view_as_imag()`
 *  [`unflatten()`][unflatten]
 *  [`unfold()`][unfold]
 *  [`unsqueeze()`][unsqueeze]
 *  [`view()`][view]
 *  [`view_as()`][view_as]
 *  [`unbind()`][unbind]
 *  [`split()`][split]
 *  `split_with_sizes()`
 *  [`swapaxes()`][swapaxes]
 *  [`swapdims()`][swapdims]
 *  [`chunk()`][chunk]
 *  [`indices()`][indices] (sparse tensor only)
 *  [`values()`][values] (sparse tensor only)

当通过索引访问张量的内容时，PyTorch 遵循 Numpy 行为，即基本索引返回视图，而高级索引返回副本。 通过基本或高级索引进行分配是就地的。 在 Numpy 索引文档中查看更多示例。

还值得一提的是一些具有特殊行为的操作：

1、reshape()、reshape\_as() 和 flatten() 可以返回视图或新张量，用户代码不应依赖于它是否为视图。

2、如果输入张量已经是连续的，则 contiguous() 返回自身，否则它通过复制数据返回一个新的连续张量。

有关 PyTorch 内部实现的更详细演练，请参阅 ezyang 关于 PyTorch 内部实现的博文。

[as_strided]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.as_strided.html#torch.Tensor.as_strided
[detach]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.detach.html#torch.Tensor.detach
[diagonal]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.diagonal.html#torch.Tensor.diagonal
[expand]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.expand.html#torch.Tensor.expand
[expand_as]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.expand_as.html#torch.Tensor.expand_as
[movedim]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.movedim.html#torch.Tensor.movedim
[narrow]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.narrow.html#torch.Tensor.narrow
[permute]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.permute.html#torch.Tensor.permute
[select]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.select.html#torch.Tensor.select
[squeeze]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.squeeze.html#torch.Tensor.squeeze
[transpose]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.transpose.html#torch.Tensor.transpose
[t]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.t.html#torch.Tensor.t
[T]: https://pytorch.org/docs/1.9.0/tensors.html#torch.Tensor.T
[real]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.real.html#torch.Tensor.real
[imag]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.imag.html#torch.Tensor.imag
[unflatten]: https://pytorch.org/docs/1.9.0/named_tensor.html#torch.Tensor.unflatten
[unfold]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.unfold.html#torch.Tensor.unfold
[unsqueeze]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.unsqueeze.html#torch.Tensor.unsqueeze
[view]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.view.html#torch.Tensor.view
[view_as]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.view_as.html#torch.Tensor.view_as
[unbind]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.unbind.html#torch.Tensor.unbind
[split]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.split.html#torch.Tensor.split
[swapaxes]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.swapaxes.html#torch.Tensor.swapaxes
[swapdims]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.swapdims.html#torch.Tensor.swapdims
[chunk]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.chunk.html#torch.Tensor.chunk
[indices]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.indices.html#torch.Tensor.indices
[values]: https://pytorch.org/docs/1.9.0/generated/torch.Tensor.values.html#torch.Tensor.values