>|h4HddlZddlZddlZddlZddlZddlZddlZddlm Z ddlm Z ddl m Z ddl mZddlmZmZ ddlmZda dd Zd Zdd Zdd Ze d ddZddZe d ddZdZddZddZddZ dZ!ddZ"y#e$r ddlmZY\wxYw)N)io)color)jit)urlparse)download_url_to_file HASH_REGEX)get_dir)_get_torch_homec ||}||}||}||}||zdz} | |zdz} tj||ftj} t|D]z} t|D]j}|t j t j |dz| z ||zz ddz t j | dz| z ||zz ddz z z| | |<l||r| tj| z } | S)N?dtypeg@)npemptyfloat32rangemathexppowsum)sizesigma amplitude normalizewidthheight sigma_horz sigma_vert mean_horz mean_vertcenter_xcenter_ygaussijs S/var/www/html/test/engine/venv/lib/python3.12/site-packages/face_alignment/utils.py _gaussianr)s,  } ~  5 3&H6!C'H HHfe_BJJ 7E 6]pu pA#dhh1q58;KU";$%&2(*-2.04!a%(:Jz\bOb9cef0gjm0m2n0o'ppE!HQK ppu % Lc Jtj|dd|zz tj|dd|zz g}tj|dd|zztj|dd|zzg}|d|jdkDs%|d|jdkDs|ddks|ddkr|Sd|zdz}tt |}|ant}t t d|d t t|d|jdt t d|dz t t d|d zg}t t d|d t t|d|jdt t d|dz t t d|d zg}t t d|dt t|d|jdg} t t d|dt t|d|jdg} |ddkDr|ddkDsJ|| ddz | d| ddz | df||ddz |d|ddz |dfz|| ddz | d| ddz | df<d||dkD<|S)Nrr)rfloorshape gauss_kernelr)intmaxmin) imagepointrulbrrgg_xg_yimg_ximg_ys r( draw_gaussianr=0s **U1XE ) *DJJuQx!e)7K,L MB **U1XE ) *DJJuQx!e)7K,L MB 1 A"Q%%++a."8BqEAIAQR u9q=D dO  s1r!uf~ C1u{{1~$> ?#c!RPQUmBT TWZ[^_`cefgchbh[iWj j kC s1r!uf~ C1u{{1~$> ?#c!RPQUmBT TWZ[^_`cefgchbh[iWj j kC Q1 SA A%?!@ AE Q1 SA A%?!@ AE FQJ3q6A:& %E!HqLq)58a<a+@@AAc!fqjQTUVQWFWY\]^Y_bcYcdghidjYjFjDkk %(Q,uQx qAeAh!6 6 E%!) Lr*chtjd}|d|d<|d|d<d|z}tjd}||z |d<||z |d<||d |z dzz|d<||d |z dzz|d <|rtj|}tj||dd }|j S) aGenerate and affine transformation matrix. Given a set of points, a center, a scale and a targer resolution, the function generates and affine transformation matrix. If invert is ``True`` it will produce the inverse transformation. Arguments: point {torch.tensor} -- the input 2D point center {torch.tensor or numpy.array} -- the center around which to perform the transformations scale {float} -- the scale of the face/object resolution {float} -- the output resolution Keyword Arguments: invert {bool} -- define wherever the function should produce the direct or the inverse transformation matrix (default: {False}) r,rri@rrrrr rrrrr)torchoneseyeinversematmulr1 r5centerscale resolutioninvert_ptht new_points r( transformrRHs" **Q-C 1XCF 1XCF  A ! A1nAdG1nAdGVAYJNS01AdGVAYJNS01AdG MM! a%q+I ==?r*cb tddg|||d}t||g|||d}|jdkDrmtj|d|dz |d|dz |jdgtj }tj |tj}n^tj|d|dz |d|dz gtj}tj |tj}|jd}|jd} tjtd|d dzt|d| |dz gtj } tjtd|d dzt|d||dz gtj } tjtd|ddzt|d| gtj } tjtd|ddzt|d|gtj } || ddz | d| ddz | dddf|| ddz | d| ddz | df<tj|t|t|ftj}|S)aCenter crops an image or set of heatmaps Arguments: image {numpy.array} -- an rgb image center {numpy.array} -- the center of the object, usually the same as of the bounding box scale {float} -- scale of the face Keyword Arguments: resolution {float} -- the size of the output cropped image (default: {256.0}) Returns: [type] -- [description] rTrrr N)dsize interpolation)rRndimrarrayr/int32zerosuint8r1r2r3cv2resize INTER_LINEAR)r4rJrKrLr6r7newDimnewImghtwdnewXnewYoldXoldYs r(croprflskU Aq665*d ;B J +VUJ MB zzA~2a52a5="Q%"Q%- ;;q>+24((<&12a52a5="Q%"Q%-8G&1 QB QB 88 QA SA^be34BHH FD 88 QA SA^be34BHH FD 88SBqEAI&BqE27rxx HD 88SBqEAI&BqE27rxx HDT!Wq[a($q'A+d1g*=q@A 47Q;tAw Q! DG 3 3 ZZs:J&H&)&6&68F Mr*T)nopythonctjd}|d|d<|d|d<d|z}tjd}||z |d<||z |d<||d |z dzz|d<||d |z dzz|d <|r2tjtjj |}tj ||dd }|jtjS) aGenerate and affine transformation matrix. Given a set of points, a center, a scale and a targer resolution, the function generates and affine transformation matrix. If invert is ``True`` it will produce the inverse transformation. Arguments: point {numpy.array} -- the input 2D point center {numpy.array} -- the center around which to perform the transformations scale {float} -- the scale of the face/object resolution {float} -- the output resolution Keyword Arguments: invert {bool} -- define wherever the function should produce the direct or the inverse transformation matrix (default: {False}) r,rrr?r@rAr rBrCr) rrErFascontiguousarraylinalgpinvdotastyperXrIs r( transform_nprns$ ''!*C 1XCF 1XCF  A q A1nAdG1nAdGVAYJNS01AdGVAYJNS01AdG  !2 3q#q#I   BHH %%r*c"|j\}}}}|j||||z}tj|d}tj|tj |ddj d} t||||\} } | | | fS)aObtain (x,y) coordinates given a set of N heatmaps. If the center and the scale is provided the function will return the points also in the original coordinate frame. Arguments: hm {torch.tensor} -- the predicted heatmaps, of shape [B, N, W, H] Keyword Arguments: center {torch.tensor} -- the center of the bounding box (default: {None}) scale {float} -- face scale (default: {None}) )axis)r/reshaperargmaxtake_along_axis expand_dimssqueeze_get_preds_fromhm) hmrJrKBCHW hm_reshapeidxscorespreds preds_origs r(get_preds_fromhmrsJAq!QAq!a%(J ))JR (C    BNN3R,Hr R Z Z[] ^F)"c65AE: *f $$r*c |j\}}}}|dz }|jdj||djtj }|dddddfdz |zdz|dddddf<t j |dddddfdz |z dz|dddddf<t|D]} t|D]} || | ddf} t|| | dfdz t|| | dfdz } } | dkDs;| dksA| dkDsG| dksMt j| | | dzf| | | dz fz | | dz| f| | dz | fz g}|| | fxxt j|dzz cc<|dz}t j|}|;|9t|D]+} t|D]} t|| | f|||d|| | f<-||fS) aObtain (x,y) coordinates given a set of N heatmaps and the coresponding locations of the maximums. If the center and the scale is provided the function will return the points also in the original coordinate frame. Arguments: hm {torch.tensor} -- the predicted heatmaps, of shape [B, N, W, H] Keyword Arguments: center {torch.tensor} -- the center of the bounding box (default: {None}) scale {float} -- face scale (default: {None}) rrNr??r T) r/repeatrrrmrrr.rr1rWsign zeros_likern)rxr~rJrKryrzr{r|rr&r'hm_pXpYdiffrs r(rwrwsJAq!Q1HC JJqM ! !!Q * 1 1"** =EAq!Gnq(A-1E!Q'NXXuQ1W~1Q67!;E!Q'N 1X4q 4AQ1W+Cq!Qw(1,c%1a..AA.EBAv"r'b1fbxxR!V_s2rAv:6a_s262:689ad rwwt}t33  44 SLEu%J e/q 9A1X 9#/!Q$K4$9 1a4  9 9 * r*c btj|jddddftj}t |jdD]J}t dD]:}t |||fdz||||dd}t |||f|dzd|||f<<Ltj|S) NrD@r rF)rMr) rrYr/rrrRr=rDtensor)target_landmarkscentersscalesheatmapsr&plandmark_cropped_coors r(create_target_heatmaprsxx)//2BB?rzzRH 8>>!$ %Yr YA$-.>q!t.Dq.H'RS*V\]^V_aclq$r !*8AqD>;PST;TVWXHQTN YY << !!r*c|jdddjd\}}|jdddjd\}}|dz}|dddf|dddfz |z}|dddf|dddfz |z}|dddfxx|zcc<|dddfxx|z cc<|dddfxx|zcc<|dddfxx|z cc<tj||gdS)a gets a batch of landmarks and calculates a bounding box that includes all the landmarks per set of landmarks in the batch :param target_landmarks: batch of landmarks of dim (n x 68 x 2). Where n is the batch size :param expansion_factor: expands the bounding box by this factor. For example, a `expansion_factor` of 0.2 leads to 20% increase in width and height of the boxes :return: a batch of bounding boxes of dim (n x 4) where the second dim is (x1,y1,x2,y2) rprrr)dimNr)rrr3r2rDcat)rexpansion_factorx_y_min_x_y_maxbb_expansion_xbb_expansion_ys r(create_bounding_boxrs"))"b!488Q8?JGQ!))"b!488Q8?JGQadmgadm37GGNadmgadm37GGN AqDM^#M AqDM^#M AqDM^#M AqDM^#M 99gw'Q //r*c^|gd}|jdk(r ||df}|S|dd|df}|S)a&Shuffle the points left-right according to the axis of symmetry of the object. Arguments: parts {torch.tensor} -- a 3D or 4D object containing the heatmaps. Keyword Arguments: pairs {list of integers} -- [order of the flipped points] (default: {None}) N)D r-r,rrr#"! -,+*/.'&%$)(6543210;:987rr>=<CBAr,.) ndimension)partspairss r( shuffle_lrrsO })  QeSj! Lam$ Lr*ctj|stj|}|r-t|j |j dz }|S|j |j dz }|S)aFlip an image or a set of heatmaps left-right Arguments: tensor {numpy.array or torch.tensor} -- [the input image or heatmaps] Keyword Arguments: is_label {bool} -- [denote wherever the input is an image or a set of heatmaps ] (default: {False}) r)rD is_tensor from_numpyrflipr)ris_labels r(rr1sp ??6 "!!&)F#(():):)>+ & 4;;CMN !!(+A()!''!*tKS+{XN +  77ell "   sD(( E1EE) r,rrFNNNNr r )F)gp@)NN)g)N)NTFN)#rrrrDrrrr[skimagerrnumbar urllib.parser torch.hubrrr BaseExceptionr r0r)r=rRrfrnrrwrrrrrrr*r(rs  !65! AEAD40!H%Pd!&!&H%*d&&R"0.2(4_ 545sB B! B!