o .wi! @sZddlZddlmZmZddlZddlmZmZddlmZm Z ddl m Z ddl m Z ddlmZded ed ed ed eeeeff d dZdedeeeedffd efddZdeded efddZdeded efddZded eded eeeeffddZded ed ed ed ef ddZ   d$ded ed eed ed ee d!d ef d"d#ZdS)%N)OptionalUnion)Tensortensor)conv2dpad)Literal)_check_same_shape)reducepredstarget hp_filter window_sizereturnc Cs|j|jkr ||j}t|||jdvr#td|jd|jdt|jdkr4|d}|d}|dks@td|d||d ksN||dkrbtd |d |d d |dd|t j }|t j }|d d d d fj|j|j d}|||fS)aUpdate and returns variables required to compute Spatial Correlation Coefficient. Args: preds: Predicted tensor target: Ground truth tensor hp_filter: High-pass filter tensor window_size: Local window size integer Return: Tuple of (preds, target, hp_filter) tensors Raises: ValueError: If ``preds`` and ``target`` have different number of channels If ``preds`` and ``target`` have different shapes If ``preds`` and ``target`` have invalid shapes If ``window_size`` is not a positive integer If ``window_size`` is greater than the size of the image )zExpected `preds` and `target` to have batch of colored images with BxCxHxW shape or batch of grayscale images of BxHxW shape. Got preds: z and target: .rrz5Expected `window_size` to be a positive integer. Got z[Expected `window_size` to be less than or equal to the size of the image. Got window_size: z and image size: xN)dtypedevice) rtor ndim ValueErrorshapelen unsqueezesizetorchfloat32r)r r r rr!^/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/torchmetrics/functional/image/scc.py _scc_updates<          r# input_imgr.cCst|tr ||||f}t|dkrtdt||ddddddd|dfjdgd}|dddddd|d dfjdgd}tj|||gdd}|ddddd|d ddfjd gd}|dddd|d dddfjd gd}tj|||gd dS) zlApplies symmetric padding to the 2D image tensor input using ``reflect`` mode (d c b a | a b c d | d c b a).rz+Expected padding to have length 4, but got Nrr)dimsrdimr) isinstanceintrrfliprcat)r$rleft_pad right_padpaddedtop_pad bottom_padr!r!r"_symmetric_reflect_pad_2dLs   .0.0r1kernelcCst|ddd}t|ddd}t|ddd}t|ddd}t|||||fd}|ddg}t||dddS)zHApplies 2D signal convolution to the input tensor with the given kernel.rrr)rrstridepadding)mathfloorrceilr1r*r)r$r2 left_padding right_padding top_paddingbottom_paddingr.r!r!r"_signal_convolve_2d\sr=cCst||dS)zOApplies 2-D Laplace filter to the input tensor with the given high pass filter.g@)r=)r$r2r!r!r"_hp_2d_laplacianhsr>windowc Cst|ddd}t|ddd}t|||||f}t|||||f}t||ddd}t||ddd}t|d|ddd|d}t|d|ddd|d}t|||ddd||} ||| fS)zrEsqrt)r r r rrrr?preds_hp target_hprBrCrDdenidxsccr!r!r"_scc_per_channel_computes       rMmean reduction)rOnoneNcsdurtgdgdgdg|durd}|dvr"td|t\fddtd D}|dkrOtjtj|d d gd d S|d kr^ttj|d d d dSdS)aCompute Spatial Correlation Coefficient (SCC_). Args: preds: predicted images of shape ``(N,C,H,W)`` or ``(N,H,W)``. target: ground truth images of shape ``(N,C,H,W)`` or ``(N,H,W)``. hp_filter: High-pass filter tensor. default: tensor([[-1,-1,-1],[-1,8,-1],[-1,-1,-1]]) window_size: Local window size integer. default: 8, reduction: Reduction method for output tensor. If ``None`` or ``"none"``, returns a tensor with the per sample results. default: ``"mean"``. Return: Tensor with scc score Example: >>> from torch import randn >>> from torchmetrics.functional.image import spatial_correlation_coefficient as scc >>> x = randn(5, 3, 16, 16) >>> scc(x, x) tensor(1.) >>> x = randn(5, 16, 16) >>> scc(x, x) tensor(1.) >>> x = randn(5, 3, 16, 16) >>> y = randn(5, 3, 16, 16) >>> scc(x, y, reduction="none") tensor([0.0223, 0.0256, 0.0616, 0.0159, 0.0170]) N)rRrR)rRrNrRrQ)rOrQz3Expected reduction to be 'mean' or 'none', but got c sVg|]'}tdd|ddddfddd|ddddfdqS)Nr)rMr).0ir r r rr!r" s Dz3spatial_correlation_coefficient..rr&)rrrrOelementwise_mean)rP) rrr#rangerrrOr+r )r r r rrP per_channelr!rUr"spatial_correlation_coefficients# rZ)NrNrO)r6typingrrrrrtorch.nn.functionalrrtyping_extensionsrtorchmetrics.utilities.checksr "torchmetrics.utilities.distributedr r)tupler#r1r=r>rErMrZr!r!r!r"s<    (&2 $(