o oi$@sUddlZddlmZmZmZmZmZddlZddlm Z ddl m Z m Z m Z ddlmZddlmZgdgdgdgZeeeed <gd gdgd gZeeeed <  d#d ed ededede fdeejdeejde fddZejddfde deee fdede fdee dee de f ddZddejdddfdeeeeefdeee fdeee fdede fd eeeefdeejdeejde fd!d"ZdS)$N)CallableListOptionalTupleUnion)Tensor) KORNIA_CHECKKORNIA_CHECK_IS_TENSORKORNIA_CHECK_SHAPE)normalize_min_max)filter2d)?r )rrrdefault_diamond_kernel)rr rdefault_square_kernel replicateswidthheight random_fn.devicedtypereturnc Cs2t|dkp|dkdd}|dkrd}||}}t|ddko"|dkd||d||g||d}||dd|ddg||d|d d d dd d df<|d d d dd d df|d d d ddd dfjdd d |d ddd df<|dkr|d ddddf|d ddddf|d ddddf|d ddddfd |d ddddf<|d}|d|d||d<|d|d||d<|d ddd df} |d ddd df|d ddd df| |d ddd df<|d ddd df|d ddd df| |d ddd df<|d}|r|dd}|S)aGenerate the diamond square image seee. Args: replicates: the num of batched replicas for the image. width: the expected image width. height: the expected image height. random_fn: the random function to generate the image seed. device: the torch device where to create the image seed. dtype: the torch dtype where to create the image seed. Return: the generated image seed of size Bx1xHxW. z#Height or Width must be equal to 3.FTz+Height must be odd and height bigger than 2rr.N)dimg@r).rr).rr).rr).rr).r ).rr).rr g@)rsum transpose) rrrrrrr"restmp1tmp2r&Q/home/ubuntu/.local/lib/python3.10/site-packages/kornia/contrib/diamond_square.py_diamond_square_seed"s. 4PN<< r(img random_scalediamond_kernel square_kernelcCst|gd|durttg|}|durttg|}|j\}}}}tj|dd|ddd|ddg|j|j d} || ddddddddddf<d} t | } | | dddddddf<| | dddddddf<| | dddddddf<| | dddddddf<|| |j|j d|} t | |} | dk|j }| d|| ||| } t | || }|dk|j }| || d|||} | S) aDoubles the image resolution by applying a single diamond square steps. Recursive application of this method creates plasma fractals. Attention! The function is differentiable and gradients are computed as well. If this function is run in the usual sense, it is more efficient if it is run in a no_grad() Args: img: a 4D tensor where dimensions are Batch, Channel, Width, Height. Width and Height must both be 2^N+1 and Batch and Channels should in the usual case be 1. random_scale: a float number in [0,1] controlling the randomness created pixels get. I the usual case, it is halved at every application of this function. random_fn: the random function to generate the image seed. diamond_kernel: the 3x3 kernel to perform the diamond step. square_kernel: the 3x3 kernel to perform the square step. Return: A tensor on the same device as img with the same channels as img and width, height of 2^(N+1)+1. )BCHWNrrrgUUUUUU?rr ) r rrtorshapetorchzerosrr ones_likesizer )r)r*rr+r,batch_sz_rrnew_imgfactorpad_compencate random_imgdiamond_regionsdiamond_centerssquare_regionssquare_centersr&r&r'_one_diamond_one_square[s0*$  rAg?g? output_size roughnessnormalize_rangecCsHtt|dkdt|ts*t|gggg||}||d|ddddg}n t||dddd}||d|dddg}|gd}t|tskt|gggg||}||d|ddddg}n|dddd}||d|dddg}|gd}|dd\}}d} |ddD]} | | 9} qd t t |dd} d t t |dd} t t t | ddt | dd} | dd | d}| dd | d}|t| |||||}|}t| D] }||}t|||}q|d d|d|f}||}|dur"|\}}t|||}|S) a8Generate Plasma Fractal Images using the diamond square algorithm. See: https://en.wikipedia.org/wiki/Diamond-square_algorithm Args: output_size: a tuple of integers with the BxCxHxW of the image to be generated. roughness: the scale value to apply at each recursion step. random_scale: the initial value of the scale for recursion. random_fn: the callable function to use to sample a random tensor. normalize_range: whether to normalize using min-max the output map. In case of a range is specified, min-max norm is applied between the provided range. device: the torch device to place the output map. dtype: the torch dtype to place the output map. Returns: A tensor with shape :math:`(B,C,H,W)` containing the fractal image. zoutput_size must be (B,C,H,W)rrr )r rrrrNr.)rlen isinstancerr1expandr viewreshapemathceillog2intminr(rangerAr contiguous)rBrCr*rrDrrrr num_samplesxp2_width p2_heightrecursion_depth seed_width seed_heightr)scaler8min_valmax_valr&r&r'diamond_squaresD     *   r\)NN)rKtypingrrrrrr3 kornia.corerkornia.core.checkrr r kornia.enhancer kornia.filtersr rfloat__annotations__rrNrrr(randrAr\r&r&r&r's    &&  <   F