PyTorch之图像和Tensor填充

在PyTorch中可以对图像和Tensor进行填充，如常量值填充，镜像填充和复制填充等。在图像预处理阶段设置图像边界填充的方式如下：

import vision.torchvision.transforms as transforms

img_to_pad = transforms.Compose([
             transforms.Pad(padding=2, padding_mode='symmetric'),
             transforms.ToTensor(),
         ])

对Tensor进行填充的方式如下：

import torch.nn.functional as F

feature = feature.unsqueeze(0).unsqueeze(0)
avg_feature = F.pad(feature, pad = [1, 1, 1, 1], mode='replicate')

这里需要注意一点的是，transforms.Pad只能对PIL图像格式进行填充，而F.pad可以对Tensor进行填充，目前F.pad不支持对2D Tensor进行填充，可以通过unsqueeze扩展为4D Tensor进行填充。

F.pad的部分源码如下：

@torch._jit_internal.weak_script
def pad(input, pad, mode='constant', value=0):
    # type: (Tensor, List[int], str, float) -> Tensor
    r"""Pads tensor.

    Pading size:
        The number of dimensions to pad is :math:`\left\lfloor\frac{\text{len(pad)}}{2}\right\rfloor`
        and the dimensions that get padded begins with the last dimension and moves forward.
        For example, to pad the last dimension of the input tensor, then `pad` has form
        `(padLeft, padRight)`; to pad the last 2 dimensions of the input tensor, then use
        `(padLeft, padRight, padTop, padBottom)`; to pad the last 3 dimensions, use
        `(padLeft, padRight, padTop, padBottom, padFront, padBack)`.

    Padding mode:
        See :class:`torch.nn.ConstantPad2d`, :class:`torch.nn.ReflectionPad2d`, and
        :class:`torch.nn.ReplicationPad2d` for concrete examples on how each of the
        padding modes works. Constant padding is implemented for arbitrary dimensions.
        Replicate padding is implemented for padding the last 3 dimensions of 5D input
        tensor, or the last 2 dimensions of 4D input tensor, or the last dimension of
        3D input tensor. Reflect padding is only implemented for padding the last 2
        dimensions of 4D input tensor, or the last dimension of 3D input tensor.

    .. include:: cuda_deterministic_backward.rst

    Args:
        input (Tensor): `Nd` tensor
        pad (tuple): m-elem tuple, where :math:`\frac{m}{2} \leq` input dimensions and :math:`m` is even.
        mode: 'constant', 'reflect' or 'replicate'. Default: 'constant'
        value: fill value for 'constant' padding. Default: 0

    Examples::

        >>> t4d = torch.empty(3, 3, 4, 2)
        >>> p1d = (1, 1) # pad last dim by 1 on each side
        >>> out = F.pad(t4d, p1d, "constant", 0)  # effectively zero padding
        >>> print(out.data.size())
        torch.Size([3, 3, 4, 4])
        >>> p2d = (1, 1, 2, 2) # pad last dim by (1, 1) and 2nd to last by (2, 2)
        >>> out = F.pad(t4d, p2d, "constant", 0)
        >>> print(out.data.size())
        torch.Size([3, 3, 8, 4])
        >>> t4d = torch.empty(3, 3, 4, 2)
        >>> p3d = (0, 1, 2, 1, 3, 3) # pad by (0, 1), (2, 1), and (3, 3)
        >>> out = F.pad(t4d, p3d, "constant", 0)
        >>> print(out.data.size())
        torch.Size([3, 9, 7, 3])

    """
    assert len(pad) % 2 == 0, 'Padding length must be divisible by 2'
    assert len(pad) // 2 <= input.dim(), 'Padding length too large'
    if mode == 'constant':
        ret = _VF.constant_pad_nd(input, pad, value)
    else:
        assert value == 0, 'Padding mode "{}"" doesn\'t take in value argument'.format(mode)
        if input.dim() == 3:
            assert len(pad) == 2, '3D tensors expect 2 values for padding'
            if mode == 'reflect':
                ret = torch._C._nn.reflection_pad1d(input, pad)
            elif mode == 'replicate':
                ret = torch._C._nn.replication_pad1d(input, pad)
            else:
                ret = input  # TODO: remove this when jit raise supports control flow
                raise NotImplementedError

        elif input.dim() == 4:
            assert len(pad) == 4, '4D tensors expect 4 values for padding'
            if mode == 'reflect':
                ret = torch._C._nn.reflection_pad2d(input, pad)
            elif mode == 'replicate':
                ret = torch._C._nn.replication_pad2d(input, pad)
            else:
                ret = input  # TODO: remove this when jit raise supports control flow
                raise NotImplementedError

        elif input.dim() == 5:
            assert len(pad) == 6, '5D tensors expect 6 values for padding'
            if mode == 'reflect':
                ret = input  # TODO: remove this when jit raise supports control flow
                raise NotImplementedError
            elif mode == 'replicate':
                ret = torch._C._nn.replication_pad3d(input, pad)
            else:
                ret = input  # TODO: remove this when jit raise supports control flow
                raise NotImplementedError
        else:
            ret = input  # TODO: remove this when jit raise supports control flow
            raise NotImplementedError("Only 3D, 4D, 5D padding with non-constant padding are supported for now")
    return ret

如果您觉得我的文章对您有所帮助，欢迎扫码进行赞赏！

 

1. Source code for torch.nn.functional

2. torch.nn.functional