3j-(V%SSKJrJr SSKrSSKJs Jr SSKJr SSKJ r SSK J r J r SSK JrJrJr SSKJr SS KJrJrJr 0r\\\4\S 'S \S '"S S\R45r"SS\R45r"SS\R45r"SS\R45rg))DictOptionalN)nn)pi)KORNIA_CHECK_LAFKORNIA_CHECK_SHAPE)SpatialGradientget_gaussian_discrete_kernel1dget_gaussian_kernel2d)rad2deg)extract_patches_from_pyramidget_laf_orientationset_laf_orientationurlszEhttps://github.com/ducha-aiki/affnet/raw/master/pretrained/OriNet.pthorinetcj\rSrSrSrS\R S\R S\R 4SjrSrg) PassLAF#zSDummy module to use instead of local feature orientation or affine shape estimator.lafimgreturncU$)zRun forward. Args: laf: :math:`(B, N, 2, 3)` img: :math:`(B, 1, H, W)` Returns: LAF, unchanged :math:`(B, N, 2, 3)` )selfrrs T/home/wildlama/miniconda3/lib/python3.13/site-packages/kornia/feature/orientation.pyforwardPassLAF.forward&s  rN) __name__ __module__ __qualname____firstlineno____doc__torchTensorr__static_attributes__rrrrr#s+] 5<< ell u|| rrc ^\rSrSrSrS S\S\S\SS4U4SjjjrS\4S jr S \ RS\ R4S jr S r U=r$) PatchDominantGradientOrientation4aModule, which estimates the dominant gradient orientation of the given patches, in radians. Zero angle points towards right. Args: patch_size: size of the (square) input patch. num_angular_bins: number of histogram bins. eps: for safe division, and arctan. patch_sizenum_angular_binsepsrNc >[TU]5 XlX l[ SS5UlX0l[R"SSSSSSS9Ul [R"5 [SS5URRSS&SSS5 [UR5S - n[URUR4XD4S 5Ulg!,(df  NP=f) Nsobelr Fcircular) kernel_sizepaddingbias padding_modeg?g@T)super__init__r+ num_ang_binsr gradientr-rConv1dangular_smoothr%no_gradr weightfloatr weighting)rr+r,r-sigma __class__s rr8)PatchDominantGradientOrientation.__init__@s $,'3  ii1!QUakl ]]_,J1c,RD   & &q )T__-3./QTYSacgh_s &$C C"cURRSURSURSURS3$)N (patch_size=z, num_ang_bins=z, eps=))rBr r+r9r-rs r__repr__)PatchDominantGradientOrientation.__repr__LsE~~&&'|DOO3DOTXTeTeSfflmqmumulvvw x rpatchc [U/SQ5 UR5up#pEX@R:wdXPR:wdUS:wa6[SURSURSUR535eURR UR 5R UR5UlURR UR 5R UR5UlURU5nUSS2SS2S4nUSS2SS2S4n[R"Xw-X--UR-5UR-n [R"XUR-5S[--n [UR 5U S [---S[-- n [R""U 5n X- n XR -n U S-UR -nS U - U -nX-n /n[%SUR 5Hln[&R("U U:HR UR 5U-UU:HR UR 5U --S 5nUR+U5 Mn [R,"US5R/S SUR 5nURU5R/S UR 5nUR1S5unnUR U-S- UR -nUS-UR -n[R2"USUR5S S55R5S 5nUn[R2"USUR5S S55R5S 5nS UU- -UU-SU-- - nS[-UR UR 5U--[UR 5- [- *nU$) aRun forward. Args: patch: :math:`(B, 1, H, W)` Returns: angle in radians: :math:`(B)` )B1HWr z#input shape should be must be [Bx1xxz]. Got Nrg@g?)r r g?)rsizer+ TypeErrorr@todtypedevicer<r:r%sqrtr-atan2rr?r9floorrangeFadaptive_avg_pool2dappendcatviewmaxgatherreshape)rrJ_CHrPrOgradsgxgymagorio_bigbo0_bigwo1_bigbo1_bigwo0_big ang_bins_listi ang_bins_iang_binsvaluesindices indices_left indices_rightleftcenterrightc_subpixangles rr(PatchDominantGradientOrientation.forwardQsl 5"67jjl q  a??&:a5doo5FaGXX_`e`j`j`l_mn **5;;7::5<<H"1144U[[ADDU\\R"mmE2 Aq> Aq>!JJrw'8488'CDt~~U!KKM:S2XEd''(C#(N;sRxH++e$/---Q;$"3"33=C'- q$++,A..A!!%++.8GqL;L;LU[[;Y\c;ccekJ   , - 99]A.33B4;L;LM&&x055b$:K:KL",,q/))G3a74;L;LL  1(9(99 ||Ha)=)=b!)DEMMbQ Xq-*?*?A*FGOOPRS$,'4%<#,+FG8wzz%++6ABU4K\K\E]]accd r)r<r-r:r9r+r@) $:0yE>)r r!r"r#r$intr?r8strrHr%r&rr' __classcell__rBs@rr)r)4sa  i3 is ie i_c i i # 1U\\1ell11rr)c^\rSrSrSrS S\S\SS4U4Sjjjr\SS\ RS\S\ R4S jj5r S \ RS\ R4S jr S r U=r$)OriNetapNetwork, which estimates the canonical orientation of the given 32x32 patches, in radians. Zero angle points towards right. This is based on the original code from paper "Repeatability Is Not Enough: Learning Discriminative Affine Regions via Discriminability"". See :cite:`AffNet2018` for more details. Args: pretrained: Download and set pretrained weights to the model. eps: to avoid division by zero in atan2. Returns: Angle in radians. Shape: - Input: (B, 1, 32, 32) - Output: (B) Examples: >>> input = torch.rand(16, 1, 32, 32) >>> orinet = OriNet() >>> angle = orinet(input) # 16 pretrainedr-rNc>[TU]5 [R"[R"SSSSSS9[R "SSS9[R "5[R"SSSSSSS9[R "SSS9[R "5[R"SSSS SSS9[R "SSS9[R "5[R"SSSSSSS9[R "SSS9[R "5[R"SS SS SSS9[R "S SS9[R "5[R"S S SSSSS9[R "S SS9[R "5[R"S 5[R"S S S SSS S9[R"5[R"S55Ul X l U(aL[RR[S[R "S5S9nUR#USSS9 UR%5 g)Nr F)r3r4r5)affine)r3strider4r5r~r1@g?Trcpu) map_location state_dict)strict)r7r8r SequentialConv2d BatchNorm2dReLUDropoutTanhAdaptiveAvgPool2dfeaturesr-r%hubload_state_dict_from_urlrrWload_state_dicteval)rrr-pretrained_dictrBs rr8OriNet.__init__s  IIaAE B NN2e , GGI IIb"!Aqu M NN2e , GGI IIb"!Aqu M NN2e , GGI IIb"!Aqu M NN2e , GGI IIb"!Aqu M NN2e , GGI IIb"!Aqu M NN2e , GGI JJt  IIb!1ad K GGI   #-  0 #ii@@h^c^j^jkp^q@rO  !>u M rrQc~[R"USSS9up#XR5- UR5U-- $)zNormalize the input by batch.)rRT)dimkeepdim)r%std_meandetach)rQr-spmps r_normalize_inputOriNet._normalize_inputs8|TBIIKBIIK#$566rrJcURURU55RSS5n[R"USS2S4S-USS2S4UR -5nU$)rLrRr1Nrrr )rrr`r%rYr-)rrJxyr|s rrOriNet.forwards\]]4007 8 = =b! D Bq!tHtOR1X-@A r)r-r)Fr)gư>)r r!r"r#r$boolr?r8 staticmethodr%r&rrr'rrs@rrrsy04etB7ELL7u7 77 U\\ ell  rrc ^\rSrSrSrSS\S\S\\RSS4U4Sjjjr S\ 4S jr S \ RS \ RS\ R4S jrS rU=r$) LAFOrienteraModule, which extracts patches using input images and local affine frames (LAFs). Then runs :class:`~kornia.feature.PatchDominantGradientOrientation` or :class:`~kornia.feature.OriNet` on patches and then rotates the LAFs by the estimated angles Args: patch_size: num_angular_bins: angle_detector: Patch orientation estimator, e.g. :class:`~kornia.feature.PatchDominantGradientOrientation` or OriNet. Nr+r,angle_detectorrc>[TU]5 XlX lU Uc&[ URUR5UlgX0lg)N)r7r8r+r9r)r)rr+r,rrBs rr8LAFOrienter.__init__sD $,  !"B4??TXTeTe"fD "0 rchURRSURSURS3$)NrEz, angle_detector=rF)rBr r+rrGs rrHLAFOrienter.__repr__s5..))*,t6GGXY]YlYlXmmnoorrrc[U5 [U/SQ5 URS5URS5:wa/[SURS5SURS535eURSSup4[ X!UR 5RSSUR UR 5nURU5RX45n[U5RU5n[U[U5U-5nU$) z}Run forward. Args: laf: :math:`(B, N, 2, 3)` img: :math:`(B, 1, H, W)` Returns: LAF_out: :math:`(B, N, 2, 3)` )rMCrOrPrz2Batch size of laf and img should be the same. Got z, Nr1rRr ) rrrS ValueErrorshaperr+r`rrview_asrr ) rrrrMNpatchesangles_radians prev_anglelaf_outs rrLAFOrienter.forwards 3 45 88A;#((1+ %QRURZRZ[\R]Q^^`adaiaijkal`mno oyy!} rs$"! Baa#WWd38nXXbii"NryyNbORYYOd3"))3r