j.ddlZddlZddlZddlmZddlmZddlm Z ddl m Z ddl m Zddl mZgd ZGd d ejZGd d ejjZGddejjZGddejjZGddejZGddejjZGddejjZdS)N)Image)gaussian_filter)pad) functional) transforms) random_shadow)Resize GaussianNoiseChannelShuffleRandomHorizontalFlip RandomShadow RandomResize GaussianBlurc eZdZdZejjddfdeeeefzde de ddffd Z d d e j d e jdzde j ee j e jfzffd Zdefd ZxZS)r a(Resize the input image to the given size >>> import torch >>> from doctr.transforms import Resize >>> transfo = Resize((64, 64), preserve_aspect_ratio=True, symmetric_pad=True) >>> out = transfo(torch.rand((3, 64, 64))) Args: size: output size in pixels, either a tuple (height, width) or a single integer for square images interpolation: interpolation mode to use for resizing, default is bilinear preserve_aspect_ratio: whether to preserve the aspect ratio of the image, if True, the image will be resized to fit within the target size while maintaining its aspect ratio symmetric_pad: whether to symmetrically pad the image to the target size, if True, the image will be padded equally on both sides to fit the target size Fsizepreserve_aspect_ratio symmetric_padreturnNctt|ttfr|n||f|d||_||_dS)NT antialias)super__init__ isinstancelisttuplerr)selfr interpolationrr __class__s c/var/www/html/Carbon-Document/venv/lib/python3.11/site-packages/doctr/transforms/modules/pytorch.pyrzResize.__init__.sW D4-!@!@RtTlTamqrrr%:"*imgtargetc|jd|jdz }|jd|jdz }|jr||krF|#t||fSt|S||kr9|jdt t |jd|z df}n8t t |jd|zd|jdf}tj|||j d}|jdd}t|jttfrd|jd|jdz d|jd|jdz f}|j rjtj|ddz tj|ddz f}|d|d|dz |d|d|dz f}t!||}||j r,|d|jdz |d|jdz f} |jr|jddd kr|j rp| d|ddddgf|dz|jdz z|ddddgf<| d|ddddgf|dz|jdz z|ddddgf<n|ddddgfxx|d|jdz zcc<|ddddgfxx|d|jdz zcc<n|jddd kr|j rW| d|d |dz|jdz z|d <| d|d |dz|jdz z|d <nX|d xx|d|jdz zcc<|d xx|d|jdz zcc<nt#d |t%j|ddfS|S)NrTrr)r+r.r).r&z=Boxes should be in the format (n_boxes, 4, 2) or (n_boxes, 4))rshaperrforwardmaxintFresizerrrrrmathceilrAssertionErrornpclip) rr#r$ target_ratio actual_ratiotmp_size raw_shape_padhalf_padoffsetr s r!r.zResize.forward9sJ y|dil2 y}sy}4 )3 ll.J.J!wws++V3377??3'' 'l** IaL#c$)A,2M.N.NPQ*R*RSDIaL<$? @ @!DDdiPQlS(3$*<MMMC "##I$)eT]33 %49Q<#)B-7DIaL39UW= TRSWW_`aWbMbcD#tnn!%V%a[39R=8(1+ RT :UUF-n|ABB'4//-O06q F111q!f9>> import torch >>> from doctr.transforms import GaussianNoise >>> transfo = GaussianNoise(0., 1.) >>> out = transfo(torch.rand((3, 224, 224))) Args: mean : mean of the gaussian distribution std : std of the gaussian distribution ?meanstdrNcdt||_||_dSrJ)rrr[rZ)rrZr[r s r!rzGaussianNoise.__init__s+  r"xc|jd|jztj|j|jzz|jz }|jtjkrL|d|zz dd tjS|| |jz ddS)Nr)devicerdtyper&) rZr[rQrandr-r_rbuint8roundclampto)rr]noises r!r.zGaussianNoise.forwards AL5:agah+O+O+OOORVRZZ 7ek ! !e O**,,221c::==EK=PP Pqw///66q!<< >> import torch >>> from doctr.transforms import GaussianBlur >>> transfo = GaussianBlur(sigma=(0.0, 1.0)) Args: sigma : standard deviation range for the gaussian kernel sigmarNcVt||_dSrJ)rr sigma_range)rrnr s r!rzGaussianBlur.__init__s'  r"r]ctjdj|j}tjt ||dd|j|j }|S)Nr&reflectg@)rnmodetruncate)rbr_) rQemptyuniform_rpitemtensorrnumpyrbr_)rr]rnblurreds r!r.zGaussianBlur.forwardsy' A')9:??AA,       '8    r") rFrKrLrMrrlrrQrRr.rUrVs@r!rrs!eE5L1!d!!!!!!%,r"rcFeZdZdZfdZdejdejfdZxZS)r z/Randomly shuffle channel order of a given imagecHtdSrJ)rr)rr s r!rzChannelShuffle.__init__s r"r#rcttj|jd}||S)Nr)rQrcr-argsort)rr# chan_orders r!r.zChannelShuffle.forwards-Z ! --5577 :r") rFrKrLrMrrQrRr.rUrVs@r!r r sd995<ELr"r cjeZdZdZdejezdejde ejezejffdZ dS)r z*Randomly flip the input image horizontallyr#r$rctjd|jkritj|}|}|jdddkrd|ddddgfz |dddddf<nd|dz |d<||fS||fS)Nr&r*rrr,)rQrcpr1hflipcopyr-)rr#r$_img_targets r!r.zRandomHorizontalFlip.forwards :a==46 ! !73<>> import torch >>> from doctr.transforms import RandomShadow >>> transfo = RandomShadow((0., 1.)) >>> out = transfo(torch.rand((3, 64, 64))) Args: opacity_range : minimum and maximum opacity of the shade N opacity_rangerctt|tr|nd|_dS)N)g?g?)rrrrr)rrr s r!rzRandomShadow.__init__s; .8.N.N^]]T^r"r]c |jtjkr}dt|tjdz |jzddtjSt||jddS#t$r|cYSwxYw)Nr`rarr&) rbrQrdr rgfloat32rrer7 ValueError)rr]s r!__call__zRandomShadow.__call__s w%+%%'DDu}D55; . UWWT!S\\RekR** %Q(:;;@@AFFF   HHH sBB=(B== C  C cd|jS)Nzopacity_range=)rrjs r!rkzRandomShadow.extra_reprs4 2444r"rJ) rFrKrLrMrrlrrQrRrrTrkrUrVs@r!rrs  __eE5L&9D&@_D______%,5<*5C55555555r"rc eZdZdZ ddeeefdeezdeezded d f fd Zd ej d e j d eej e j ffdZ d e fdZxZS)raRandomly resize the input image and align corresponding targets >>> import torch >>> from doctr.transforms import RandomResize >>> transfo = RandomResize((0.3, 0.9), preserve_aspect_ratio=True, symmetric_pad=True, p=0.5) >>> out = transfo(torch.rand((3, 64, 64))) Args: scale_range: range of the resizing factor for width and height (independently) preserve_aspect_ratio: whether to preserve the aspect ratio of the image, given a float value, the aspect ratio will be preserved with this probability symmetric_pad: whether to symmetrically pad the image, given a float value, the symmetric padding will be applied with this probability p: probability to apply the transformation g333333?g?F? scale_rangerrrrNct||_||_||_||_t |_dSrJ)rrrrrrr _resize)rrrrrr s r!rzRandomResize.__init__sE &%:"* r"r#r$c tjd|jkr%tjj|j}tjj|j}t|jd|zt|jd|zf}| |t|j tr|j n)ttjd|j kt|j tr|j n)ttjd|j k||\}}||fS||fS)Nr&r'r()rr)rQrcrr6randomuniformrr0r-rrrrPr)rr#r$scale_hscale_wnew_sizerrs r!r.zRandomResize.forward!s% :a==46 ! !i')9:Gi')9:GCIbMG344c#)B-':Q6R6RSHDLLd8$??'?d&@&@%*Q--4+==>>d0$77?d00%*Q--4+==>>)6MD'= F{r"cFd|jd|jd|jd|jS)Nz scale_range=rArBz, p=)rrrrrjs r!rkzRandomResize.extra_repr4s`Ud.UUHbUUtxuGUUMQMSUU Ur")rFFr)rFrKrLrMrrlrPrrQrRr6rSr.rTrkrUrVs@r!rrs$,6.3&+  5%<(  $e| e|          5<elTVT^F^@_&UCUUUUUUUUr"r)r3ryr6rQ PIL.Imager scipy.ndimagertorch.nn.functionalrtorchvision.transformsrr1rTr __all__r nnModuler rr r rrrr"r!rs   ))))))######222222222222&&&&&&   ^5^5^5^5^5QX^5^5^5B33333EHO333<58?B     UX_        11    &5&5&5&5&558?&5&5&5R3U3U3U3U3U58?3U3U3U3U3Ur"