3j[2%SSKJrJrJrJrJr SSKrSSKJs J r SSKJr SSK J r SSK JrJr SSKJr SSKJr Sr\\S 'S S /r\\\S '/S Qr\\\S'/SQr\\\S'\\\S.r\\\\4\S'SVs0sH nUSUS3_M snr\\\4\S'S0S\S\\\R<44SjjrS\S\S\R<4Sjr "SS\RB5r""SS \RB5r#"S!S"\RB5r$S#\S$\\\R<4S\R<4S%jr%"S&S'\RB5r&"S(S)\RB5r'"S*S+\RB5r(S#\S,\S\\\44S-jr)"S.S/\RB5r*gs snf)1)AnyDictListTupleUnionN)nn)pi)GaussianBlur2dSpatialGradient)cart2pol)create_meshgridg;f?sqrt2gpu?g "? COEFFS_N1_K1)"g#[?x@ٔ+?K '"? COEFFS_N2_K8)rrrgˈfF? COEFFS_N3_K8)xyrhophithetaCOEFFS)cartpolarconcatzChttps://github.com/manyids2/mkd_pytorch/raw/master/mkd_pytorch/mkd-z-64.pthurls patch_sizereturncr[XSS9nUSSS2SS2S4nUSSS2SS2S4n[X#5upEX#XES.nU$)z1Get cartesian and polar parametrizations of grid.Theightwidthnormalized_coordinatesrN)xyrhophi)r r )rkgridr%r&r'r( grid_dicts L/home/wildlama/miniconda3/lib/python3.13/site-packages/kornia/feature/mkd.py get_grid_dictr,*sN :X\ ]E aAqjA aAqjA~HC8I d1d2c [R"[R"U5[R"U5SS9up#[R"X#/SS9R SS5nUR [R 5$)z&Get order for doing kronecker product.ij)indexingdim)torchmeshgridarangestackreshapetoint64)r.r/grid_d1grid_d2 kron_orders r+get_kron_orderrA4s]~~ell2&6 R8HSWXGg/Q7??AFJ == %%r-cx^\rSrSrSrS U4SjjrS\RS\R4SjrS\ 4Sjr Sr U=r $) MKDGradients;a[nn.Module, which computes gradients of given patches, stacked as [magnitudes, orientations]. Given gradients $g_x$, $g_y$ with respect to $x$, $y$ respectively, - $\mathbox{mags} = $\sqrt{g_x^2 + g_y^2 + eps}$ - $\mathbox{oris} = $\mbox{torch.tan}^{-1}(\nicefrac{g_y}{g_x})$. Args: patch_size: Input patch size in pixels. Returns: gradients of given patches. Shape: - Input: (B, 1, patch_size, patch_size) - Output: (B, 2, patch_size, patch_size) Example: >>> patches = torch.rand(23, 1, 32, 32) >>> gradient = MKDGradients() >>> g = gradient(patches) # 23x2x32x32 rcP>[TU]5 SUl[SSSS9Ulg)N:0yE>diffr$F)modeorder normalized)super__init__epsr grad)self __class__s r+rLMKDGradients.__init__Ss% #qUK r-r%c[U[R5(d[S[ U535e[ UR 5S:Xd[SUR 35eURU5*nUSS2SS2SSS2SS24nUSS2SS2SSS2SS24n[R"[X4UR5SS9nU$)N&Input type is not a torch.Tensor. Got -Invalid input shape, we expect Bx1xHxW. Got: rr$r4) isinstancer7Tensor TypeErrortypelenshape ValueErrorrNcatr rM)rOr%grads_xygxgyr&s r+forwardMKDGradients.forwardYs!U\\**DT!WINO O177|q LQWWIVW WIIaL= aAq!m $ aAq!m $ IIhrtxx0a 8r-c.URR$N)rP__name__rOs r+__repr__MKDGradients.__repr__es~~&&&r-)rMrNrN) re __module__ __qualname____firstlineno____doc__rLr7rWrastrrg__static_attributes__ __classcell__rPs@r+rCrC;s:.L  %,, '#''r-rCc ^\rSrSrSrS\S\\\\\4\ \\\4S44SS4U4Sjjr S \ RS\ R4S jr S\4S jrS rU=r$) VonMisesKerneliann.Module, which computes parameters of Von Mises kernel given coefficients, and embeds given patches. Args: patch_size: Input patch size in pixels. coeffs: List of coefficients. Some examples are hardcoded in COEFFS, Returns: Von Mises embedding of given parametrization. Shape: - Input: (B, 1, patch_size, patch_size) - Output: (B, d, patch_size, patch_size) Examples: >>> oris = torch.rand(23, 1, 32, 32) >>> vm = VonMisesKernel(patch_size=32, ... coeffs=[0.14343168, ... 0.268285, ... 0.21979234]) >>> emb = vm(oris) # 23x7x32x32 rcoeffs.rNc>[TU]5 Xl[R"U5nUR SU5 [ U5S- nX@lSU-S-Ul[R"SSX/5n[R"U5S-nURSSS5n[R"SU-S-/5n[R"U5USUS-&[R"USS5XtS-S&URSSS5nUR SU5 UR SU5 UR SU5 g)Nrur$r3r6emb0frangeweights)rKrLrr7tensorregister_bufferrZndonesr9r;zerossqrt) rOrrub_coeffsr|rwrxryrPs r+rLVonMisesKernel.__init__s $<<' Xx0 K!OQzz1a89a1$Aq)++q1uqyk* ::h/!a% ::hqrl3A//"a+ VT* Xv. Y0r-r%c[U[R5(d[S[ U535e[ UR 5S:XaUR SS:wa[SUR 35e[UR[R5(d[S[ U535eURRU5RURS5SSS5nURRU5U-n[R"U5n[R"U5n[R"X$U/SS9nUR U-nU$)NrSrTr$rUz%Emb0 type is not a torch.Tensor. Got rr4)rVr7rWrXrYrZr[r\rwr<repeatsizerxcossinr]ry)rOr%rwrxemb1emb2 embeddings r+raVonMisesKernel.forwards!U\\**DT!WINO O177|q AGGAJ!OLQWWIVW W$))U\\22CDG9MN Nyy||A%%affQiAq9"Q&yy yy IIt40a8 LL9, r-c URRSURSURSURSUR S3 $)N (patch_size=z, n=z, d=z , coeffs=))rPrerr|r}rurfs r+rgVonMisesKernel.__repr__sN..))*,t6GtDFF8SWX\X^X^W__himitithuuvwwr-)r}r|r)rerjrkrlrmintrrfloatrrLr7rWrarnrgrorprqs@r+rsrsis.131d5;L6MuUZ[`be[eUfhkUkOl6l0m1rv10%,,$x#xxr-rsc^\rSrSrSrSS\S\SS4U4SjjjrS\RS\R4S jr S \RS\R4S jr S\ 4S jr S rU=r$)EmbedGradientsann.Module that computes gradient embedding, weighted by sqrt of magnitudes of given patches. Args: patch_size: Input patch size in pixels. relative: absolute or relative gradients. Returns: Gradient embedding. Shape: - Input: (B, 2, patch_size, patch_size) - Output: (B, 7, patch_size, patch_size) Examples: >>> grads = torch.rand(23, 2, 32, 32) >>> emb_grads = EmbedGradients(patch_size=32, ... relative=False) >>> emb = emb_grads(grads) # 23x7x32x32 rrelativerNc>[TU]5 XlX lSUl[ U[ SS9Ul[XSS9n[USS2SS2SS2S4USS2SS2SS2S45upEURSU5 g) NrFrrruTr rr$r() rKrLrrrMrsrkernelr r r{)rOrrr)_r(rPs r+rLEmbedGradients.__init__sz $ % 6'?S  z\`a%1a +U1aA:->? UC(r-magscJ[R"XR-5nU$)z@Embed square roots of magnitudes with eps for numerical reasons.)r7rrM)rOrs r+emb_magsEmbedGradients.emb_magsszz$/* r-gradsc[U[R5(d[S[ U535e[ UR 5S:Xd[SUR 35eUSS2SS2SS2SS24nUSS2SS2SS2SS24nUR(aX0RRU5- nURU5URU5-nU$)NrSrTz-Invalid input shape, we expect Bx2xHxW. Got: r$) rVr7rWrXrYrZr[r\rr(r<rr)rOrrorisr&s r+raEmbedGradients.forwards%..DT%[MRS S5;;1$LU[[MZ[ [QAq[!QAq[! ==((++d++D KK  d 3 3r-chURRSURSURS3$)Nrz , relative=r)rPrerrrfs r+rgEmbedGradients.__repr__s4..))*,t6G{SWS`S`Raabccr-)rMrrr) F)rerjrkrlrmrboolrLr7rWrrarnrgrorprqs@r+rrsq* )3 )t ) ) )U\\ell U\\ ell d#ddr-r kernel_typegridsc4S[[- [S- [S- S.nUS:XaSnSS/nO US:XaS nS S /n[UR55nXS RS nUR 5VVs0sH upxXxX'-_M n nnU R 5VVs0sH3upxXxR S 5R S 5R5_M5 n nn[U[WS9n [U[US9n U "U WS 5R5n U "XS5R5n [U RU R5nU RS USS2S 45U RS USS2S45-nU$s snnfs snnf)z>> emb_ori = torch.rand(23, 7, 32, 32) >>> ese = ExplicitSpacialEncoding(kernel_type='polar', ... fmap_size=32, ... in_dims=7, ... do_gmask=True, ... do_l2=True) >>> desc = ese(emb_ori) # 23x175x32x32 r fmap_sizein_dimsdo_gmaskdo_l2rNc>[T U]5 US;a[US35eXlX lX0lX@lXPl[U5Ul SUl [URUR5nUR (a"URSS9Ul X`R-nURSURS55 URSUlUR UR -UlUR"UlUR'5upxURSU5 URSU5 g) N)rrz" is not valid, use polar or cart).r)sigmaembrridx1)rKrLNotImplementedErrorrrrrrr,gridgmaskr get_gmaskr{rr[d_embout_dimsodims init_kron) rOrrrrrrrrrPs r+rL ExplicitSpacialEncoding.__init__#s  / /% 5W&XY Y&"   !),  't'7'7C ==c2DJ "C UCMM!$45))A, !\\DJJ6 -- ^^%  VT* VT*r-rcURSURSR5- n[R"SUS--US-- 5nU$)zCompute Gaussian mask.r'r6r3)rmaxr7exp)rOrnorm_rhors r+r!ExplicitSpacialEncoding.get_gmaskKsL99U#dii&6&:&:&<< "x{*UAX56 r-c[URUR5n[RR [R UR5n[R"USUSS2S45nX1SS2S44$)z3Initialize helper variables to calculate kronecker.r$Nr) rArrr7jitannotaterWrr)rOkron_embrs r+r!ExplicitSpacialEncoding.init_kronQsadllDJJ7yy!!%,,9!!$41:6!Q$Zr-r%cV[U[R5(d[S[ U535e[ UR 5S:HUR SUR:H-(d%[SURSUR 35e[RR[RUR5n[R"USU5nX0R-nURSS9nUR(a[ R""USS9nU$)NrSrTr$z!Invalid input shape, we expect Bxz xHxW. Got: )r3r4)rVr7rWrXrYrZr[rr\rrrrrsumrF normalize)rOr%rroutputs r+raExplicitSpacialEncoding.forwardXs!U\\**DT!WINO OQWW"qwwqzT\\'AB@kZ[ZaZaYbcd dyy!!%,, :!!!Q- !' ::[[Q/F r-cURRSURSURSURSUR SUR SURS3$)Nz (kernel_type=z , fmap_size= , in_dims=z , out_dims=z , do_gmask=z, do_l2=r)rPrerrrrrrrfs r+rg ExplicitSpacialEncoding.__repr__esv~~&&'(++,-()||n% ' 'ZZL  # r-) rrrrrrrrrr)rrTT)rerjrkrlrmrnrrrLrr7rWrrrrargrorprqs@r+rrs:# &+&+&+ &+  &+  &+ &+&+Pu  5u||!;<   %,,   #   r-rc2^\rSrSrSrSS\S\\\\\\R44S4S\ S\ S\ S \ S S4U4S jjjr S\\\\\R44S S4S jr SS jrSSjrSSjrSSjrS\RS \R4SjrS \4SjrSrU=r$) Whiteningiqann.Module, performs supervised or unsupervised whitening. This is based on the paper "Understanding and Improving Kernel Local Descriptors". See :cite:`mukundan2019understanding` for more details. Args: xform: Variant of whitening to use. None, 'lw', 'pca', 'pcaws', 'pcawt'. whitening_model: Dictionary with keys 'mean', 'eigvecs', 'eigvals' holding Tensors. in_dims: Dimensionality of input descriptors. output_dims: (int) Dimensionality reduction. keval: Shrinkage parameter. t: Attenuation parameter. Returns: l2-normalized, whitened descriptors. Shape: - Input: (B, in_dims, fmap_size, fmap_size) - Output: (B, out_dims, fmap_size, fmap_size) Examples: >>> descs = torch.rand(23, 238) >>> whitening_model = {'pca': {'mean': torch.zeros(238), ... 'eigvecs': torch.eye(238), ... 'eigvals': torch.ones(238)}} >>> whitening = Whitening(xform='pcawt', ... whitening_model=whitening_model, ... in_dims=238, ... output_dims=128, ... keval=40, ... t=0.7) >>> wdescs = whitening(descs) # 23x128 xformwhitening_modelNr output_dimskevaltrc>[TU]5 XlX0lXPlX`lSUl[XC5nX@l[R"[R"U5SS9Ul [R"[R"U5SS2SU24SS9Ul[R"[R "U5SUSS9UlUbUR%U5 gg)NrT) requires_grad)rKrLrrrrpvalminrr Parameterr7rmeaneyeevecsr~evalsload_whitening_parameters)rOrrrrrrrPs r+rLWhitening.__init__s     +/ &LLW!5TJ \\%))G"4Q  _"EUYZ \\%**W"5l{"CSWX  &  * *? ; 'r-cpURS:XaSOSnXnUSURlUSSS2SUR24URlUSSURUR lUR URURURS.nX@R"5 g)Nlwpcareigvecseigvals)rrpcawspcawt) rrdatarrr _modify_pca _modify_lw _modify_pcaws _modify_pcawt)rOralgowh_model modificationss r+r#Whitening.load_whitening_parametersszzT)tu"(!&) "9-a1C43C3C1C.CD "9-.@0@0@A ##//''''   jj!#r-cSUlg)zModify powerlaw parameter.g?N)rrfs r+rWhitening._modify_pcas  r-cg)zNo modification required.Nrfs r+rWhitening._modify_lwsr-cURURnSU- UR-U-nUR[R"[R "US55-URlg)zShrinkage for eigenvalues.r$N)rrrr7diagpowr)rOalphars r+rWhitening._modify_pcawssS 4::&e)tzz)U2**uzz%))E42H'II r-cSUR-nUR[R"[R"UR U55-URlg)zAttenuation for eigenvalues.rN)rrr7rrrr)rOms r+rWhitening._modify_pcawts; 466M**uzz%))DJJ2J'KK r-r%c[U[R5(d[S[ U535e[ UR 5S:Xd[SUR 35eXR- nXR-n[R"U5[R"[R"U5UR5-n[R"USS9$)NrSr3z)Invalid input shape, we expect NxD. Got: r$r4)rVr7rWrXrYrZr[r\rrsignrabsrrrrOr%s r+raWhitening.forwards!U\\**DT!WINO O177|q H RS S M N JJqMEIIeiilDII> >{{1!$$r-cURRSURSURSURS3$)Nz(xform=r, output_dims=r)rPrerrrrfs r+rgWhitening.__repr__s>..))*'$**Z ~Ucdhdtdtcuuvwwr-) rrrrrrrrr)(gffffff?ri)rerjrkrlrmrnrrr7rWrrrLrrrrrrargrorprqs@r+rrqs!P<<tCc5<<.?)@$@A4GH< <  <  < < <<:$c4U\\HYCZ>Z9[$`d$"(J L %%%,,%x#xxr-rc ^\rSrSrSrSS\S\S\S\S\SS 4 U4S jjjrS \RS\R4S jr S\4S jr Sr U=r $) MKDDescriptoriaCnn.Module that computes Multiple Kernel local descriptors. This is based on the paper "Understanding and Improving Kernel Local Descriptors". See :cite:`mukundan2019understanding` for more details. Args: patch_size: Input patch size in pixels. kernel_type: Parametrization of kernel ``'concat'``, ``'cart'``, ``'polar'``. whitening: Whitening transform to apply ``None``, ``'lw'``, ``'pca'``, ``'pcawt'``, ``'pcaws'``. training_set: Set that model was trained on ``'liberty'``, ``'notredame'``, ``'yosemite'``. output_dims: Dimensionality reduction. Returns: Explicit cartesian or polar embedding. Shape: - Input: :math:`(B, in_{dims}, fmap_{size}, fmap_{size})`. - Output: :math:`(B, out_{dims}, fmap_{size}, fmap_{size})`, Examples: >>> patches = torch.rand(23, 1, 32, 32) >>> mkd = MKDDescriptor(patch_size=32, ... kernel_type='concat', ... whitening='pcawt', ... training_set='liberty', ... output_dims=128) >>> desc = mkd(patches) # 23x128 rr whitening training_setrrNc>[TU]5 XlX lX0lX@lSUS- -Ul[SUR UR 4S5Ul[5Ul SnSnURS:XaXg/O UR/Ul SUl US US 0n0Ul URHon [XU S 9n [XU R R"S 9n [$R&"X5URU 'U=RU R- sl Mq [)XPR5UlURb|[,R.R1[2UR[,R4"S 5S9n Xn [7X=URUR*S9UlUR*Ul UR;5 g)Nffffff?@r( replicaterrrrTFrrrrrcpu map_locationrr)rKrLrrr"r#rr smoothingrC gradientsparametrizationsrfeatsrrrr}r Sequentialrrr7hubload_state_dict_from_urlrdevicerwhitening_layereval)rOrrr"r#rpolar_scart_srelative_orientationsparametrizationgradient_embeddingspatial_encodingwhitening_modelsrrPs r+rLMKDDescriptor.__init__s ) +'!-JO, 'TZZ0H+V%595E5E5Q 1X\XhXhWi !($ > #44O!/:fuPv!w 6+K]KdKdKfKf  +---8J*]DJJ ' JJ*00 0J 5!$K < >> %$yyAAT%%&U\\%5H B  /> %$$Q'Ar-c URRSURSURSURSUR SUR S3 $)Nrz, kernel_type=z , whitening=z, training_set=rr)rPrerrr"r#rrfs r+rgMKDDescriptor.__repr__Oso~~&&'(//*+++,-() --./++,A  / r-) r3r1rrrr2rrr0r#r"r8)rrrlibertyr)rerjrkrlrmrrnrLr7rWrargrorprqs@r+r!r!s@# % 000 0  0  0 00du|| 6 #  r-r!r#c[RR[U[R"S5S9nX!nU$)zLoad whitening model.r,r-)r7r5r6rr7)rr#r@rs r+load_whitening_modelrLZs<yy99${:KZ_ZfZfglZm9n&4O r-c ^\rSrSrSrSS\S\S\S\S\SS 4 U4S jjjrS \RS\R4S jr S r U=r $)SimpleKDiaz+Example to write custom Kernel Descriptors.rrr"r#rrNc>>[T U]5 US:HnSUS- -nXl[SXw4S5n[ 5n [ XS9n [ X!U RRS9n [U[X$5U RUS9n [R"XXU 5Ulg) Nrr%r&r'r)r*r+r/)rKrLrr rCrrrr}rrLrrr4rF)rOrrr"r#rrrr0r1oriesewhrPs r+rLSimpleKD.__init__ds $/j2o.$"6E>;G  N  F%+]`]g]g]i]ij  +KFPSPYPYgr  iCbI r-r%c$URU5$rd)rFrs r+raSimpleKD.forward|s}}Qr-)rFr)rrrrJr) rerjrkrlrmrrnrLr7rWrarorprqs@r+rNrNas5" % JJJ J  J  J JJ0  %,,  r-rN)r)+typingrrrrrr7torch.nn.functionalr functionalrkornia.constantsr kornia.filtersr r kornia.geometry.conversionsr kornia.geometry.gridr rr__annotations__rrrrrnrrrWr,rAModulerCrsrrrrr!rLrN)rs0r+r_s$10 :00"u!'4 d5k4> d5k>J d5kJ(4 Wc!dS$u+ d) ( LQCwWW (d38n c4U\\0A+B&s&& &+'299+'\CxRYYCxL6dRYY6dr#d3 ;L6MRWR^R^>g biig Tox oxdt BIIt ncc3h ryy {s?F