o yiP@s:ddlmZmZmZmZmZddlZddlmZddlm Z ddl m Z ddl mZmZmZddlmZddlmZd ed ed eeeffd d Z        d5d ed ededeeeefdeeeefdeededededed eeeeefffddZ d6ded e d!d efd"d#Z         d7d ed ededeeeefdeeeefd e d!deededededed eeeeefffd$d%Z       d8d ed ededeeeefdeeeefdeededed&ee d'd eeeffd(d)Z       * d9d ed ededeeeefdeeeefdeededed+eeeeeeefeed,ffd&ee d'd efd-d.Z d6d/ed e d!d efd0d1Z        * 2d:d ed ededeeeefdeeeefd e d!deededed+eed,fd&ee d'd efd3d4ZdS);)ListOptionalSequenceTupleUnionN)Tensor) functional)Literal)_gaussian_kernel_2d_gaussian_kernel_3d_reflection_pad_3d)_check_same_shapereducepredstargetreturncCsR|j|jkr ||j}t||t|jdvr%td|jd|jd||fS)zUpdates and returns variables required to compute Structural Similarity Index Measure. Checks for same shape and type of the input tensors. Args: preds: Predicted tensor target: Ground truth tensor )zMExpected `preds` and `target` to have BxCxHxW or BxCxDxHxW shape. Got preds: z and target: .)dtypetor lenshape ValueError)rrrV/home/ubuntu/.local/lib/python3.10/site-packages/torchmetrics/functional/image/ssim.py_ssim_check_inputss  rT? {Gz?Q?Fgaussian_kernelsigma kernel_size data_rangek1k2return_full_imagereturn_contrast_sensitivityc #Cs\t|jdk} t|ts| rd|gnd|g}t|ts)| r$d|gnd|g}t|t|jdkrCtdt|dt|jt|dvrRtdt|t|t|jdkrltdt|dt|jt|dvr{tdt||r| rtdtd d |Drtd |d td d |Drtd|d |durt||||}t||d} t||d} |j } | d}|j }dd|D}|ddd}|ddd}| r|ddd}t ||||}t ||||}|rt ||||| }n#tj|||||fdd}tj|||||fdd}|r2t||||| }|sMtj|dg|R|| dttj||| d}t||||||||f}| ritj|||d}ntj|||d}||jd}|dd}|dd}|d|d}|d|}|d|}|d|}d||| }|||| }d|| |||| |} | r| d|| || || f}!n | d|| || f}!| r||}"|"d|| || f}"|!|!jddd|"|"jdddfS|r"|!|!jddd| fS|!|!jdddS)a#Computes Structual Similarity Index Measure. Args: preds: estimated image target: ground truth image gaussian_kernel: If true (default), a gaussian kernel is used, if false a uniform kernel is used sigma: Standard deviation of the gaussian kernel, anisotropic kernels are possible. Ignored if a uniform kernel is used kernel_size: the size of the uniform kernel, anisotropic kernels are possible. Ignored if a Gaussian kernel is used data_range: Range of the image. If ``None``, it is determined from the image (max - min) k1: Parameter of SSIM. k2: Parameter of SSIM. return_full_image: If true, the full ``ssim`` image is returned as a second argument. Mutually exlusive with ``return_contrast_sensitivity`` return_contrast_sensitivity: If true, the contrast term is returned as a second argument. The luminance term can be obtained with luminance=ssim/contrast Mutually exclusive with ``return_full_image`` rz`kernel_size` has dimension zD, but expected to be two less that target dimensionality, which is: )r+r*zMExpected `kernel_size` dimension to be 2 or 3. `kernel_size` dimensionality: zWArguments `return_full_image` and `return_contrast_sensitivity` are mutually exclusive.css$|] }|ddkp |dkVqdS)r+rNr).0xrrr js"z_ssim_update..z8Expected `kernel_size` to have odd positive number. Got rcss|]}|dkVqdS)rNr)r,yrrrr.msz.Expected `sigma` to have positive number. Got NcSs$g|]}td|dddqS)g @g?r+r0)int)r,srrr ys$z _ssim_update..rreflect)mode)rdevice)groupsr.)rr isinstancerranymaxminpowr6sizerr r Fpadr torchonesprodtensorcatconv3dconv2dsplitrreshapemean)#rrr"r#r$r%r&r'r(r)is_3dc1c2r6channelrgauss_kernel_sizepad_hpad_wpad_dkernel input_listoutputs output_list mu_pred_sq mu_target_sqmu_pred_target sigma_pred_sqsigma_target_sqsigma_pred_targetupperlowerssim_idx_full_imagessim_idxcontrast_sensitivityrrr _ssim_update.s        "      $ rbelementwise_mean similarities reduction)rcsumnoneNcC t||S)aApplies the specified reduction to pre-computed structural similarity. Args: similarities: per image similarities for a batch of images. reduction: a method to reduce metric score over individual batch scores - ``'elementwise_mean'``: takes the mean - ``'sum'``: takes the sum - ``'none'`` or ``None``: no reduction will be applied Returns: The reduced SSIM score r)rdrerrr _ssim_compute ric  CsVt||\}}t||||||||| | } t| tr$| \} } t| || fS| } t| |S)a(Computes Structual Similarity Index Measure. Args: preds: estimated image target: ground truth image gaussian_kernel: If true (default), a gaussian kernel is used, if false a uniform kernel is used sigma: Standard deviation of the gaussian kernel, anisotropic kernels are possible. Ignored if a uniform kernel is used kernel_size: the size of the uniform kernel, anisotropic kernels are possible. Ignored if a Gaussian kernel is used reduction: a method to reduce metric score over labels. - ``'elementwise_mean'``: takes the mean - ``'sum'``: takes the sum - ``'none'`` or ``None``: no reduction will be applied data_range: Range of the image. If ``None``, it is determined from the image (max - min) k1: Parameter of SSIM. k2: Parameter of SSIM. return_full_image: If true, the full ``ssim`` image is returned as a second argument. Mutually exclusive with ``return_contrast_sensitivity`` return_contrast_sensitivity: If true, the constant term is returned as a second argument. The luminance term can be obtained with luminance=ssim/contrast Mutually exclusive with ``return_full_image`` Return: Tensor with SSIM score Raises: TypeError: If ``preds`` and ``target`` don't have the same data type. ValueError: If ``preds`` and ``target`` don't have ``BxCxHxW shape``. ValueError: If the length of ``kernel_size`` or ``sigma`` is not ``2``. ValueError: If one of the elements of ``kernel_size`` is not an ``odd positive number``. ValueError: If one of the elements of ``sigma`` is not a ``positive number``. Example: >>> from torchmetrics.functional import structural_similarity_index_measure >>> preds = torch.rand([3, 3, 256, 256]) >>> target = preds * 0.75 >>> structural_similarity_index_measure(preds, target) tensor(0.9219) )rrbr9tupleri)rrr"r#r$rer%r&r'r(r)similarity_pack similarityimagerrr#structural_similarity_index_measures$<  ro normalizerelusimplec CsBt||||||||dd \} } |dkrt| } t| } | | fS)NT)r)rr)rbrArr) rrr"r#r$r%r&r'rpsimrarrr_get_normalized_sim_and_css    rugǺ?g48EG?ga4?g??g9EGr?betas.c  Cs<g} t|jdk} t|ts| rd|gnd|g}t|ts+| r&d|gnd|g}|ddt|ksC|ddt|krTtdt|ddt|dtd t|d d} |d| |d d krtdt|d |d d |d d | d|d| |d d krtdt|d |d d |d d | dtt|D]?} t||||||||| d \}}| |t|dkrt |d}t |d}qt|dkrt |d}t |d}qtd|| d<t | }| dkr|d d}t j||jddd }||}t j|d d}|S)aQComputes Multi-Scale Structual Similarity Index Measure. Adapted from: https://github.com/jorge-pessoa/pytorch-msssim/blob/master/pytorch_msssim/__init__.py. Args: preds: estimated image target: ground truth image kernel_size: size of the gaussian kernel sigma: Standard deviation of the gaussian kernel reduction: a method to reduce metric score over labels. - ``'elementwise_mean'``: takes the mean - ``'sum'``: takes the sum - ``'none'`` or ``None``: no reduction will be applied data_range: Range of the image. If ``None``, it is determined from the image (max - min) k1: Parameter of structural similarity index measure. k2: Parameter of structural similarity index measure. betas: Exponent parameters for individual similarities and contrastive sensitives returned by different image resolutions. normalize: When MultiScaleSSIM loss is used for training, it is desirable to use normalizes to improve the training stability. This `normalize` argument is out of scope of the original implementation [1], and it is adapted from https://github.com/jorge-pessoa/pytorch-msssim instead. Raises: ValueError: If the image height or width is smaller then ``2 ** len(betas)``. ValueError: If the image height is smaller than ``(kernel_size[0] - 1) * max(1, (len(betas) - 1)) ** 2``. ValueError: If the image width is smaller than ``(kernel_size[0] - 1) * max(1, (len(betas) - 1)) ** 2``. rr*r+r8z)For a given number of `betas` parameters zH, the image height and width dimensions must be larger than or equal to rr0rz and kernel size z', the image height must be larger than z&, the image width must be larger than )rp)r+r+)r+r+r+z(length of kernel_size is neither 2 nor 3rs)r6)axis)rrr9rr>rr;rangeruappendr? avg_pool2d avg_pool3drAstackrDr6viewrC)rrr"r#r$r%r&r'rwrpmcs_listrK _betas_div_rtra mcs_stack mcs_weighted mcs_per_imagerrr_multiscale_ssim_update9sZ2  0          rrcCrh)aApplies the specified reduction to pre-computed multi-scale structural similarity. Args: mcs_per_image: per image similarities for a batch of images. reduction: a method to reduce metric score over individual batch scores - ``'elementwise_mean'``: takes the mean - ``'sum'``: takes the sum - ``'none'`` or ``None``: no reduction will be applied Returns: The reduced multi-scale structural similarity r)rrerrr_multiscale_ssim_computerjrrrc Cs|t| ts tdt| trtdd| Dstd| r%| dvr%tdt||\}}t||||||||| | } t| |S)ao Computes `MultiScaleSSIM`_, Multi-scale Structual Similarity Index Measure, which is a generalization of Structual Similarity Index Measure by incorporating image details at different resolution scores. Args: preds: Predictions from model of shape ``[N, C, H, W]`` target: Ground truth values of shape ``[N, C, H, W]`` kernel_size: size of the gaussian kernel sigma: Standard deviation of the gaussian kernel reduction: a method to reduce metric score over labels. - ``'elementwise_mean'``: takes the mean - ``'sum'``: takes the sum - ``'none'`` or ``None``: no reduction will be applied data_range: Range of the image. If ``None``, it is determined from the image (max - min) k1: Parameter of structural similarity index measure. k2: Parameter of structural similarity index measure. betas: Exponent parameters for individual similarities and contrastive sensitivies returned by different image resolutions. normalize: When MultiScaleSSIM loss is used for training, it is desirable to use normalizes to improve the training stability. This `normalize` argument is out of scope of the original implementation [1], and it is adapted from https://github.com/jorge-pessoa/pytorch-msssim instead. Return: Tensor with Multi-Scale SSIM score Raises: TypeError: If ``preds`` and ``target`` don't have the same data type. ValueError: If ``preds`` and ``target`` don't have ``BxCxHxW shape``. ValueError: If the length of ``kernel_size`` or ``sigma`` is not ``2``. ValueError: If one of the elements of ``kernel_size`` is not an ``odd positive number``. ValueError: If one of the elements of ``sigma`` is not a ``positive number``. Example: >>> from torchmetrics.functional import multiscale_structural_similarity_index_measure >>> preds = torch.rand([3, 3, 256, 256], generator=torch.manual_seed(42)) >>> target = preds * 0.75 >>> multiscale_structural_similarity_index_measure(preds, target, data_range=1.0) tensor(0.9627) References: [1] Multi-Scale Structural Similarity For Image Quality Assessment by Zhou Wang, Eero P. Simoncelli and Alan C. Bovik `MultiScaleSSIM`_ z3Argument `betas` is expected to be of a type tuple.css|]}t|tVqdS)N)r9float)r,betarrrr.szAmultiscale_structural_similarity_index_measure..z5Argument `betas` is expected to be a tuple of floats.rqzNArgument `normalize` to be expected either `None` or one of 'relu' or 'simple')r9rkrallrrr) rrr"r#r$rer%r&r'rwrprrrr.multiscale_structural_similarity_index_measures >  r)TrrNr r!FF)rc) TrrrcNr r!FF)TrrNr r!N)TrrNr r!rvN) TrrrcNr r!rvrr) typingrrrrrrArtorch.nnrr?typing_extensionsr $torchmetrics.functional.image.helperr r r torchmetrics.utilities.checksr "torchmetrics.utilities.distributedrrboolrr1rbrirorurrrrrrrsp               U     k