o oi(@sddlmZddlmZddlZddlmZddlmZmZm Z m Z m Z m Z m Z GdddeZGdd d eZGd d d eZdS) ) annotations)CallableN)nn)ModuleTensor as_tensorstacktensorwhere zeros_likecsTeZdZUdZded<ded<ddfdd ZdddZdddZdddZZ S)_HausdorffERLossBasea#Base class for binary Hausdorff loss based on morphological erosion. This is an Hausdorff Distance (HD) Loss that based on morphological erosion,which provided a differentiable approximation of Hausdorff distance as stated in :cite:`karimi2019reducing`. The code is refactored on top of `here `__. Args: alpha: controls the erosion rate in each iteration. k: the number of iterations of erosion. reduction: Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the weighted mean of the output is taken, 'sum': the output will be summed. Returns: Estimated Hausdorff Loss. zCallable[..., Tensor]convmax_pool@ meanalphafloatkint reductionstrreturnNonecs0t||_||_||_|d|dS)Nkernel)super__init__rrrregister_buffer get_kernel)selfrrr __class__K/home/ubuntu/.local/lib/python3.10/site-packages/kornia/losses/hausdorff.pyr3s z_HausdorffERLossBase.__init__rcCst),Get kernel for image morphology convolution.)NotImplementedError)rr"r"r#r:sz_HausdorffERLossBase.get_kernelpredtargetcCs||d}t|j|j|jd}t||j|jd}tj||jtjd}|ddd}t |j D]L}|j |||dd} | d} d| | dk<| | } | |  } | | dk} | }|rp| | | | }t|| || } || |d|j}| }q1|S)Ndevicedtype)weightpaddinggroupsg?r)rrr*r+r torch ones_likeboolsizerangerr rsqueezeanyr r)rr&r'boundrerodedmaskr/rdilationerosion erosion_max erosion_min_to_normto_norm_erosion_to_fillr"r"r#perform_erosion>s&    z$_HausdorffERLossBase.perform_erosioncsjddjddkrddkrddks,tdjdjddkr;tdtfd d tdD}jd krX|}|Sjd krc| }|Sjd krk |St djd)aoCompute Hausdorff loss. Args: pred: predicted tensor with a shape of :math:`(B, C, H, W)` or :math:`(B, C, D, H, W)`. Each channel is as binary as: 1 -> fg, 0 -> bg. target: target tensor with a shape of :math:`(B, 1, H, W)` or :math:`(B, C, D, H, W)`. Returns: Estimated Hausdorff Loss. r(Nrr-zTPrediction and target need to be of same size, and target should not be one-hot.Got z and .zInvalid target value.c sTg|]&}dd||dft|ktdjjdtdjjdqS)Nr-r)r)rBr r r*r+).0ir&rr'r"r# ws z0_HausdorffERLossBase.forward..rsumnonez reduction `z` has not been implemented yet.) shaper4 ValueErrormaxitemrr5rrrHr%)rr&r'outr"rFr#forwardas2>     z_HausdorffERLossBase.forward)rrr)rrrrrrrrrrr&rr'rrr) __name__ __module__ __qualname____doc____annotations__rrrBrO __classcell__r"r"r r#r s   #r c<eZdZdZejZedZ d ddZ d fd d Z Z S) HausdorffERLossa>Binary Hausdorff loss based on morphological erosion. Hausdorff Distance loss measures the maximum distance of a predicted segmentation boundary to the nearest ground-truth edge pixel. For two segmentation point sets X and Y , the one-sided HD from X to Y is defined as: .. math:: hd(X,Y) = \max_{x \in X} \min_{y \in Y}||x - y||_2 Furthermore, the bidirectional HD is: .. math:: HD(X,Y) = max(hd(X, Y), hd(Y, X)) This is an Hausdorff Distance (HD) Loss that based on morphological erosion, which provided a differentiable approximation of Hausdorff distance as stated in :cite:`karimi2019reducing`. The code is refactored on top of `here `__. Args: alpha: controls the erosion rate in each iteration. k: the number of iterations of erosion. reduction: Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the weighted mean of the output is taken, 'sum': the output will be summed. Examples: >>> hdloss = HausdorffERLoss() >>> input = torch.randn(5, 3, 20, 20) >>> target = (torch.rand(5, 1, 20, 20) * 2).long() >>> res = hdloss(input, target) r-rrcCs,tgdgdgdgg}|d}|dS)r$rr-rr-r-r-g?N)r )rcrossrr"r"r#rszHausdorffERLoss.get_kernelr&r'cs|dkrtd|d||dkr%|dkr%|jtjks:td|dd|d|d t ||S) a9Compute Hausdorff loss. Args: pred: predicted tensor with a shape of :math:`(B, C, H, W)`. Each channel is as binary as: 1 -> fg, 0 -> bg. target: target tensor with a shape of :math:`(B, 1, H, W)`. Returns: Estimated Hausdorff Loss. zOnly 2D images supported. Got rCr-rz+Expect long type target value in range (0, z). (z, )) dimrKrLr4minr+r1longrrOrr&r'r r"r#rOs *$zHausdorffERLoss.forwardrPrQ) rRrSrTrUr1conv2dr rAdaptiveMaxPool2drrrOrWr"r"r r#rYs $  rYcrX) HausdorffERLoss3DaMBinary 3D Hausdorff loss based on morphological erosion. Hausdorff Distance loss measures the maximum distance of a predicted segmentation boundary to the nearest ground-truth edge pixel. For two segmentation point sets X and Y , the one-sided HD from X to Y is defined as: .. math:: hd(X,Y) = \max_{x \in X} \min_{y \in Y}||x - y||_2 Furthermore, the bidirectional HD is: .. math:: HD(X,Y) = max(hd(X, Y), hd(Y, X)) This is a 3D Hausdorff Distance (HD) Loss that based on morphological erosion, which provided a differentiable approximation of Hausdorff distance as stated in :cite:`karimi2019reducing`. The code is refactored on top of `here `__. Args: alpha: controls the erosion rate in each iteration. k: the number of iterations of erosion. reduction: Specifies the reduction to apply to the output: 'none' | 'mean' | 'sum'. 'none': no reduction will be applied, 'mean': the weighted mean of the output is taken, 'sum': the output will be summed. Examples: >>> hdloss = HausdorffERLoss3D() >>> input = torch.randn(5, 3, 20, 20, 20) >>> target = (torch.rand(5, 1, 20, 20, 20) * 2).long() >>> res = hdloss(input, target) r-rrcCsTtgdgdgdgg}tgdgdgdgg}t|||gdd}|dS)r$rZr[)rrrr-g$I$I?N)r r)rr\r8rr"r"r#rszHausdorffERLoss3D.get_kernelr&r'cs.|dkrtd|dt||S)aBCompute 3D Hausdorff loss. Args: pred: predicted tensor with a shape of :math:`(B, C, D, H, W)`. Each channel is as binary as: 1 -> fg, 0 -> bg. target: target tensor with a shape of :math:`(B, 1, D, H, W)`. Returns: Estimated Hausdorff Loss. zOnly 3D images supported. Got rC)r_rKrrOrbr r"r#rOs zHausdorffERLoss3D.forwardrPrQ) rRrSrTrUr1conv3dr rAdaptiveMaxPool3drrrOrWr"r"r r#res $   re) __future__rtypingrr1r kornia.corerrrrr r r r rYrer"r"r"r#s   $tD