a d5 @sddlmZmZmZmZddlZddlmZddlmm Z ddl m Z eee e fee e e e ffZee ee e ffee ee e ffee e e e fdddZGdddejZGd d d ejZdejee ee e ffee ee e ffee ee e ffee efejd d d Zejee dfee dfejdddZdejee ee e ffee ee e ffee efejdddZdS))OptionalTupleUnioncastN)_pair) original_size window_sizereturncCs~ttttft|}ttttft|}tttttfddd}||d|d\}}||d|d\}}||||fS)aCompute required padding to ensure chaining of :func:`extract_tensor_patches` and :func:`combine_tensor_patches` produces expected result. Args: original_size: the size of the original tensor. window_size: the size of the sliding window used while extracting patches. Return: The required padding for `(top, bottom, left, right)` as a tuple of 4 ints. Example: >>> image = torch.arange(12).view(1, 1, 4, 3) >>> padding = compute_padding((4,3), (3,3)) >>> out = extract_tensor_patches(image, window_size=(3, 3), stride=(3, 3), padding=padding) >>> combine_tensor_patches(out, original_size=(4, 3), window_size=(3, 3), stride=(3, 3), unpadding=padding) tensor([[[[ 0, 1, 2], [ 3, 4, 5], [ 6, 7, 8], [ 9, 10, 11]]]]) .. note:: This function is supposed to be used in conjunction with :func:`extract_tensor_patches` and :func:`combine_tensor_patches`. )dim1dim2r cSsL||dkrd}d}n.|||}|ddkr<|d}}n|}d}||fS)Nr)r r p1p2tmpr r g/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/kornia/contrib/extract_patches.pypaddim)s   zcompute_padding..paddimr)rrintr)rrrZpadtZpadbZpadlZpadrr r rcompute_padding s  rcsreZdZdZd eeeeeffeeeeeeffeeeefddfdd Z e j e j dd d Z Z S) ExtractTensorPatchesaModule that extract patches from tensors and stack them. In the simplest case, the output value of the operator with input size :math:`(B, C, H, W)` is :math:`(B, N, C, H_{out}, W_{out})`. where - :math:`B` is the batch size. - :math:`N` denotes the total number of extracted patches stacked in - :math:`C` denotes the number of input channels. - :math:`H`, :math:`W` the input height and width of the input in pixels. - :math:`H_{out}`, :math:`W_{out}` denote to denote to the patch size defined in the function signature. left-right and top-bottom order. * :attr:`window_size` is the size of the sliding window and controls the shape of the output tensor and defines the shape of the output patch. * :attr:`stride` controls the stride to apply to the sliding window and regulates the overlapping between the extracted patches. * :attr:`padding` controls the amount of implicit zeros-paddings on both sizes at each dimension. The parameters :attr:`window_size`, :attr:`stride` and :attr:`padding` can be either: - a single ``int`` -- in which case the same value is used for the height and width dimension. - a ``tuple`` of two ints -- in which case, the first `int` is used for the height dimension, and the second `int` for the width dimension. :attr:`padding` can also be a ``tuple`` of four ints -- in which case, the first two ints are for the height dimension while the last two ints are for the width dimension. Args: window_size: the size of the sliding window and the output patch size. stride: stride of the sliding window. padding: Zero-padding added to both side of the input. Shape: - Input: :math:`(B, C, H, W)` - Output: :math:`(B, N, C, H_{out}, W_{out})` Returns: the tensor with the extracted patches. Examples: >>> input = torch.arange(9.).view(1, 1, 3, 3) >>> patches = extract_tensor_patches(input, (2, 3)) >>> input tensor([[[[0., 1., 2.], [3., 4., 5.], [6., 7., 8.]]]]) >>> patches[:, -1] tensor([[[[3., 4., 5.], [6., 7., 8.]]]]) rrN)rstridepaddingr cs,tt||_t||_t||_dSN)super__init__rrrr)selfrrr __class__r rrvs   zExtractTensorPatches.__init__inputr cCst||j|j|jdS)N)rr)extract_tensor_patchesrrrrr r r rforwardszExtractTensorPatches.forward)rr)__name__ __module__ __qualname____doc__rrrrPadTypertorchTensorr# __classcell__r r rrr<s< rcsjeZdZdZd eeeeeffeeeeeffeeefddfdd Ze j e j ddd Z Z S) CombineTensorPatchesag Module that combine patches from tensors. In the simplest case, the output value of the operator with input size :math:`(B, N, C, H_{out}, W_{out})` is :math:`(B, C, H, W)`. where - :math:`B` is the batch size. - :math:`N` denotes the total number of extracted patches stacked in - :math:`C` denotes the number of input channels. - :math:`H`, :math:`W` the input height and width of the input in pixels. - :math:`H_{out}`, :math:`W_{out}` denote to denote to the patch size defined in the function signature. left-right and top-bottom order. * :attr:`original_size` is the size of the original image prior to extracting tensor patches and defines the shape of the output patch. * :attr:`window_size` is the size of the sliding window used while extracting tensor patches. * :attr:`unpadding` is the amount of padding to be removed. This value must be the same as padding used while extracting tensor patches. The parameters :attr:`original_size`, :attr:`window_size`, and :attr:`unpadding` can be either: - a single ``int`` -- in which case the same value is used for the height and width dimension. - a ``tuple`` of two ints -- in which case, the first `int` is used for the height dimension, and the second `int` for the width dimension. :attr:`unpadding` can also be a ``tuple`` of four ints -- in which case, the first two ints are for the height dimension while the last two ints are for the width dimension. Args: patches: patched tensor. original_size: the size of the original tensor and the output patch size. window_size: the size of the sliding window used. unpadding: remove the padding added to both side of the input. Shape: - Input: :math:`(B, N, C, H_{out}, W_{out})` - Output: :math:`(B, C, H, W)` Example: >>> out = extract_tensor_patches(torch.arange(16).view(1, 1, 4, 4), window_size=(2, 2), stride=(2, 2)) >>> combine_tensor_patches(out, original_size=(4, 4), window_size=(2, 2), stride=(2, 2)) tensor([[[[ 0, 1, 2, 3], [ 4, 5, 6, 7], [ 8, 9, 10, 11], [12, 13, 14, 15]]]]) .. note:: This function is supposed to be used in conjunction with :class:`ExtractTensorPatches`. rN)rr unpaddingr cs,tt||_t||_t||_dSr)rrrrrr-)rrrr-rr rrs   zCombineTensorPatches.__init__rcCst||j|j|j|jdS)N)rr-)combine_tensor_patchesrrr-r"r r rr#szCombineTensorPatches.forward)r) r$r%r&r'rrrr(rr)r*r#r+r r rrr,s;  r,)patchesrrrr-r c CsTt|jdkrtd|jttttft|}ttttft|}ttttft|}|d|dks|d|dkrtd|d|d|rttt|}t|dvrt d t|d krt|d}t|d}n|d d }|d d }ttttttf||}|d|d |d |ddk}|d|d|d|ddk}|rj|srtd |d|d |d |d|d|d|d|df}|j d|d|dg|jdd R} t t j | |ddddd} t t j | |ddddd} |rPttttttf|}t| dd|D} | S)aRestore input from patches. See :class:`~kornia.contrib.CombineTensorPatches` for details. Args: patches: patched tensor with shape :math:`(B, N, C, H_{out}, W_{out})`. original_size: the size of the original tensor and the output patch size. window_size: the size of the sliding window used while extracting patches. stride: stride of the sliding window. unpadding: remove the padding added to both side of the input. Return: The combined patches in an image tensor with shape :math:`(B, C, H, W)`. Example: >>> out = extract_tensor_patches(torch.arange(16).view(1, 1, 4, 4), window_size=(2, 2), stride=(2, 2)) >>> combine_tensor_patches(out, original_size=(4, 4), window_size=(2, 2), stride=(2, 2)) tensor([[[[ 0, 1, 2, 3], [ 4, 5, 6, 7], [ 8, 9, 10, 11], [12, 13, 14, 15]]]]) .. note:: This function is supposed to be used in conjunction with :func:`extract_tensor_patches`. z/Invalid input shape, we expect BxNxCxHxW. Got: rrz-Only stride == window_size is supported. Got z and z0.Please feel free to drop a PR to Kornia Github.r zHUnpadding must be either an int, tuple of two ints or tuple of four intsr NzInsufficient padding)dimcSsg|] }| qSr r ).0ir r r z*combine_tensor_patches..)lenshape ValueErrorrrrrNotImplementedErrorr(AssertionErrorviewr)catchunkZsqueezeFpad) r/rrrr-pad_vertpad_horzZ hpad_checkZ wpad_checkZpatches_tensorZrestored_tensorr r rr.sB!      $$ (""r..)r window_sizesstridesr cs|dd\}}td|t||D]\}}}||||}q.|jdgdfddDR}|j|d|g|RS)Nr rrc3s|]}|tVqdSr)r<)r8r6dimsr r &r;z,_extract_tensor_patchesnd..r4)sizeranger6zipZunfoldZpermute contiguousrA)r rHrIZ batch_sizeZ num_channelsr6Z patch_sizerr rJr_extract_tensor_patchesnds ,rQr)r rrrr cCst|stdt|t|jdkr:td|j|rttt |}t|dvr`t dt|dkrt |d}t |d}n|d d}|dd }tt t t t t f||}t ||}t|t |t |S) adFunction that extract patches from tensors and stack them. See :class:`~kornia.contrib.ExtractTensorPatches` for details. Args: input: tensor image where to extract the patches with shape :math:`(B, C, H, W)`. window_size: the size of the sliding window and the output patch size. stride: stride of the sliding window. padding: Zero-padding added to both side of the input. Returns: the tensor with the extracted patches with shape :math:`(B, N, C, H_{out}, W_{out})`. Examples: >>> input = torch.arange(9.).view(1, 1, 3, 3) >>> patches = extract_tensor_patches(input, (2, 3)) >>> input tensor([[[[0., 1., 2.], [3., 4., 5.], [6., 7., 8.]]]]) >>> patches[:, -1] tensor([[[[3., 4., 5.], [6., 7., 8.]]]]) z,Input input type is not a torch.Tensor. Got r2z-Invalid input shape, we expect BxCxHxW. Got: r1zFPadding must be either an int, tuple of two ints or tuple of four intsr rrN)r)Z is_tensor TypeErrortyper<r=r>rr(rr@rrrDrErQ)r rrrrFrGr r rr!*s        r!)r)rr)typingrrrrr)Ztorch.nnnnZtorch.nn.functionalZ functionalrDZtorch.nn.modules.utilsrrr(rModulerr,r*r.rQr!r r r rs>   $ 1IN  R