+ j^RIHt^RIHt^RIt^RIt^RIHt^RIH t ^RI H t ^RI H t ^RIHt^R IHtHt^R IHt^R IHtHtHtHt^R IHt^R IHt]!RR7]!RR]444t!RR]P>4t Rt!R/Rlt"RRlt#R0RRllt$!RR]P>4t%!RR]P>4t&!RR]P>4t'R R!lt(!R"R#]P>4t)!R$R%]P>4t*]!R&R']44t+R(R)lt,R*R+lt-]!R,R7!R-R.]+44t.R'R..t/R#)1)Callable) dataclassN)nn pad_sequence)ACT2FN)create_bidirectional_mask)FlashAttentionKwargs)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack) ModelOutputTransformersKwargsauto_docstringcan_return_tuple)AutoModelForKeypointDetectionLightGlueConfiga Base class for outputs of LightGlue keypoint matching models. Due to the nature of keypoint detection and matching, the number of keypoints is not fixed and can vary from image to image, which makes batching non-trivial. In the batch of images, the maximum number of matches is set as the dimension of the matches and matching scores. The mask tensor is used to indicate which values in the keypoints, matches, matching_scores and prune tensors are keypoint matching information. ) custom_introcPa]tRt^&toRtRtRtRtRtRt Rt Rt Rt V3Rlt RtVtR#)LightGlueKeypointMatchingOutputaW loss (`torch.FloatTensor` of shape `(1,)`, *optional*): Loss computed during training. matches (`torch.FloatTensor` of shape `(batch_size, 2, num_matches)`): Index of keypoint matched in the other image. matching_scores (`torch.FloatTensor` of shape `(batch_size, 2, num_matches)`): Scores of predicted matches. keypoints (`torch.FloatTensor` of shape `(batch_size, num_keypoints, 2)`): Absolute (x, y) coordinates of predicted keypoints in a given image. prune (`torch.IntTensor` of shape `(batch_size, num_keypoints)`): Pruning mask indicating which keypoints are removed and at which layer. mask (`torch.BoolTensor` of shape `(batch_size, num_keypoints)`): Mask indicating which values in matches, matching_scores, keypoints and prune are keypoint matching information. hidden_states (`Tuple[torch.FloatTensor, ...]`, *optional*): Tuple of `torch.FloatTensor` (one for the output of each stage) of shape `(batch_size, 2, num_channels, num_keypoints)` returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True` attentions (`Tuple[torch.FloatTensor, ...]`, *optional*): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, 2, num_heads, num_keypoints, num_keypoints)` returned when `output_attentions=True` is passed or when `config.output_attentions=True` Nc<V^8dQh/S[PR,;R&S[PR,;R&S[PR,;R&S[PR,;R&S[PR,;R&S[PR,;R&S[S[P,R,;R&S[S[P,R,;R &#) Nlossmatchesmatching_scores keypointsprunemask hidden_states attentions)torch FloatTensor IntTensortuple)format __classdict__s"q/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/lightglue/modeling_lightglue.py __annotate__,LightGlueKeypointMatchingOutput.__annotate__&sF   d ")GH   %,IJ&&-4KL  4'.MN ??T !(OP   d ")QR**+d29STe''(4/6U)__name__ __module__ __qualname____firstlineno____doc__rrrrrrr r!__annotate_func____static_attributes____classdictcell__r's@r(rr&s<0&*D(,G04O*.I$(E%)D59M26JUr+rcTaa]tRt^StoV3RlV3RlltRV3RlRlltRtVtV;t#)LightGluePositionalEncoderc <V^8dQhRS[/#rconfigr)r&r's"r(r)'LightGluePositionalEncoder.__annotate__Tslllr+c<\SV`4\P!^VPVP ,^,RR7VnR#)rFbiasN)super__init__rLineardescriptor_dimnum_attention_heads projectorselfr: __class__s&&r(r@#LightGluePositionalEncoder.__init__Ts: 1f&;&;v?Y?Y&Y]^&^ejkr+c <V^8dQhRS[PRS[R,RS[S[P,S[S[PS[P3,,/#)rroutput_hidden_statesNreturnr"Tensorboolr%)r&r's"r(r)r;XsM   =AD[  u|| uU\\5<<%?@ @ r+cVPV4pVP^RR7p\P!V4p\P!V4pWV3pV'dWC3pV#V3pV#)rdim)rDrepeat_interleaver"cossin)rFrrJprojected_keypoints embeddingscosinessinesoutputs&&& r(forward"LightGluePositionalEncoder.forwardXsl#nnY7(::1":E ))J' *%% 6J*2 R\P] r+)rDF r-r.r/r0r@r[r3r4 __classcell__rGr's@@r(r7r7Ss ll   r+r7c~VR,pVR,p\P!V)V.RR7PR4pV#).rP).:NNr).:NrrR)r"stackflatten)xx1x2rot_xs& r( rotate_halfrjds: 8B 9B KK"b r * 2 22 6E Lr+cbVPpVP4pVP4pVPV4pVPV4pW,\V4V,,pW,\V4V,,pVP VR7VP VR73#)aIApplies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. dtype)rmfloat unsqueezerjto)qkrTrU unsqueeze_dimrmq_embedk_embeds&&&&& r(apply_rotary_pos_embrvls$ GGE  A  A -- &C -- &Cw;q>C/0Gw;q>C/0G ::E: "GJJUJ$; ;;r+cdV^8dQhR\PR\R\P/#)rr n_reprKr"rMint)r&s"r(r)r)s. U UU\\ U# U%,, Ur+cVPwr#rEV^8XdV#VR,PW#WV4pVPW#V,WE4#)z This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) )NNNr|Nr|r|)shapeexpandreshape)r rxbatchnum_key_value_headsslenhead_dims&& r( repeat_kvrsU 2?1D1D.E z!"23::5W\dlmM  e(CT TTr+cV^8dQhR\PR\PR\PR\PR\PR,R\R\R \ \ ,/#) rmodulequerykeyvalueattention_maskNscalingdropoutkwargs)rModuler"rMrnr r)r&s"r(r)r)sw%% II% <<% % << % LL4' %  %%'(%r+c "\W P4p\W0P4p \P!WP ^^44V,p Ve W,p \ P PV R\PR7PVP4p \ P PWVPR7p \P!W4p V P ^^4P4p W3#)r)rQrm)ptrainingrR)rnum_key_value_groupsr"matmul transposer functionalsoftmaxfloat32rprmrr contiguous) rrrrrrrr key_states value_states attn_weights attn_outputs &&&&&&&, r(eager_attention_forwardrs3 ; ;$>&B\B\$\!}}d* !'!9!9ii    : :T]] JQWQfQf  ii    : :T]] JQWQfQf  ii    : :T]] JQWQfQf  ii  & & 68J8JQWQfQf  r+cf<V^8dQhRS[PRS[S[PS[P3,R,RS[PR,RS[PR,RS[PR,RS[S[,RS[S[PS[PR,3,/#) rr position_embeddingsNrencoder_hidden_statesencoder_attention_maskrrK)r"rMr%r r )r&r's"r(r)rs*)*)||*)#5<<#=>E*) t+ *) %||d2 *) !& t 3 *)-.*) u||U\\D00 1*)r+c VPRRp.VORNVPN5pVPV4PV4P ^^4p VRJp V 'dTMTp V 'dTMTp VP V 4PV4P ^^4p VP V 4PV4P ^^4pVeVwpp\WVV4wr\P!VPP\4pT!VV V VV 3RVP'gRM VPRVP/VBwppVP !.VORN5!P#4pVP%V4pVV3#)NrrrR)r}rrviewrrrrvr get_interfacer:_attn_implementationrrrrrrr)rFr rrrrr input_shape hidden_shape query_statesis_cross_attentioncurrent_statescurrent_attention_maskrrrTrUattention_interfacerrs&&&&&&, r(r[LightGlueAttention.forwards$))#2. 88b8$--8 {{=166|DNNqRST 2$>2D.-;M!7Sa[[055lCMMaQRS {{>277 EOOPQSTU  **HC';LVY[^'_ $L(?(M(M KK , ,.E) %8     " % $}}}C$2H2H % LL %  % ! \"));;;;FFH kk+. L((r+) rr:rrrrrrrrr)NNNN) r-r.r/r0r1r@r[r3r4r_r`s@@r(rrs!G  .*)*)*)r+rcPaa]tRt^toV3RlV3RlltV3RlRltRtVtV;t#) LightGlueMLPc <V^8dQhRS[/#r9r)r&r's"r(r)LightGlueMLP.__annotate__sZZZr+cv<\SV`4Wn\VP,Vn\ P!VPVP4Vn \ P!VPVP4Vn \ P!VPRR7Vn R#)T)elementwise_affineN)r?r@r:r hidden_act activation_fnrrAintermediate_sizefc1rfc2 LayerNorm layer_normrEs&&r(r@LightGlueMLP.__init__s{  #F$5$5699V55v7O7OP99V55v7I7IJ,,v'?'?TXYr+cN<V^8dQhRS[PRS[P/#)rr rKr"rM)r&r's"r(r)rs#U\\ellr+cVPV4pVPV4pVPV4pVPV4pV#N)rrrr)rFr s&&r(r[LightGlueMLP.forwardsB/  6 **=9 / r+)rr:rrrr^r`s@@r(rrs ZZr+rcTaa]tRtRtoV3RlV3RlltRV3RlRlltRtVtV;t#)LightGlueTransformerLayeric&<V^8dQhRS[RS[/#rr)r&r's"r(r)&LightGlueTransformerLayer.__annotate__s...3.r+c<\SV`4\W4Vn\ V4Vn\W4Vn\ V4VnR#r)r?r@rself_attentionrself_mlpcross_attention cross_mlprs&&&r(r@"LightGlueTransformerLayer.__init__s@ 0C$V, 1&D%f-r+c(<V^8dQhRS[PRS[PRS[PRS[R,RS[R,RS[S[PS[S[P,R,S[S[P,R,3,/#)r descriptorsrrrJNoutput_attentionsrKrL)r&r's"r(r)r sH>H>\\H><<H> H> #Tk H>  $; H> u||U5<<047u||9Lt9SS TH>r+c`V'dRMRpV'dRMRpV'd Wa3,pVPwrp VPVVVVR7wr\P!W.RR7p VP V 4pW,pV'dW3pVP R^W4P ^4P WV 4pVe7VP R^^^V 4P ^4P V^^V 4MRpVPVVVVR7wpp\P!VV.RR7pVPV4pVV,pV'dEVV3pVVP WV 43,X,VP WV 43,V,pV'dW|3,V3,pWV3#)N)rrrrP)rrrr,rR) r}rr"catrrfliprr)rFrrrrJrall_hidden_statesall_attentions batch_size num_keypointsrBattention_outputself_attentionsintermediate_states output_statesself_attention_descriptorsself_attention_hidden_statesrrcross_attention_outputcross_attentionscross_intermediate_statescross_output_statescross_attention_hidden_statess&&&&&& r(r[!LightGlueTransformerLayer.forward s#7BD0d  1N B 4?4E4E1 >-1,?,?  ))/ -@- ) $ii(GRP &9: %0%@" ,?+O ( ' . .r1m T T!W WZ ? )  " "2q!Q > C CA F N Nz[\^_an o 483G3G &"7#9/ 4H4 0 0 %*II/IKa.bhj$k!"nn-FG03FF -FH[,\ )!-55jQ_`bc./&&z.QST0 0  +.@@DTCVVN~==r+)rrrr)FFr^r`s@@r(rrs!..H>H>H>r+rcV^8dQhR\PR\PR\PR\P/#)r similarity matchability0 matchability1rKr)r&s"r(r)r)Xs<    -2\\ JO,,  \\ r+cVPwr4p\PPV4\PPV4P ^^4,p\PP V^4p\PP VP RR4P 4^4P RR4pVPW4^,V^,3^4p Wx,V,V RRV1RV13&\PPVPR4)4V RRR1R3&\PPVPR4)4V RRRR13&V #)z;create the log assignment matrix from logits and similarityr|NrRrc) r}rr logsigmoidr log_softmaxrnew_fullsqueeze) rrrrnum_keypoints_0num_keypoints_1 certaintiesscores0scores1scoress &&& r(sigmoid_log_double_softmaxrXsM4>3C3C0J--**=9BMM--k: % 2t7J7JS`SgSg(hlp(pp %--jAoq-` &t,&t,#&>&>r2&FF  << aMBDJ((,EJ((,66r2>E=D#// 5;;zGWGW;X;\;\]J((5 #++J!OQ qQ %d+%d+,JW r+cN<V^8dQhRS[PRS[P/#rrrKr)r&r's"r(r)rs#ELLU\\r+c VPV4p\PPV4P R4pV#)z0Get matchability of descriptors as a probabilityrR)r rrsigmoidr)rFrr s&& r(get_matchability.LightGlueMatchAssignmentLayer.get_matchabilitys7((5 }},,\:BB2F r+)rBr r ) r-r.r/r0r@r[r#r3r4r_r`s@@r(rrgs*II4r+rcPaa]tRtRtoV3RlV3RlltV3RlRltRtVtV;t#)LightGlueTokenConfidenceLayeric <V^8dQhRS[/#r9r)r&r's"r(r)*LightGlueTokenConfidenceLayer.__annotate__s999r+cp<\SV`4\P!VP^4VnR#)rbN)r?r@rrArBtokenrEs&&r(r@&LightGlueTokenConfidenceLayer.__init__s& YYv44a8 r+cN<V^8dQhRS[PRS[P/#r r)r&r's"r(r)r(s#5<<ELLr+cVPVP44p\PP V4P R4pV#)rbrR)r*detachrrr"r)rFrr*s&& r(r[%LightGlueTokenConfidenceLayer.forwards= ;--/0 %%e,44R8 r+)r*r^r`s@@r(r&r&s99 r+r&cHa]tRtRtoRtRtRtR tRtRt Rt V3Rlt Rt Vt R #) LightGluePreTrainedModelizz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. lightglue pixel_valuesFTc&<V^8dQh/S[;R&#r9r)r&r's"r(r)%LightGluePreTrainedModel.__annotate__s r+r,N)image)r-r.r/r0r1base_model_prefixmain_input_nameinput_modalitiessupports_gradient_checkpointing_supports_flash_attn_supports_sdpar2r3r4r5s@r(r1r1s5 $$O!&+#Nr+r1cV^8dQhR\PR\R\\P\P3,/#)rr thresholdrK)r"rMrnr%)r&s"r(r)r)s:$$ELL$U$uU\\[`[g[gMgGh$r+c|VPwppVRRR1RR13,P^4pVRRR1RR13,P^4pVPpVPp\P!VP^,VP R7R,p\P!VP^,VP R7R,p WP ^V48Hp WP ^V48Hp VPP4pVP^4p \P!WV 4p \P!WP ^V4V 4pWV8,pWP ^V4,p\P!WR4p\P!VVR4p\P!Wg.4P^^4PV^,R4p\P!W.4P^^4PV^,R4pVV3#)z1obtain matches from a score matrix [Bx M+1 x N+1]r|NrrR)r}maxindicesr"arangergathervaluesexp new_tensorwhererdrr)rr>r_max0max1matches0matches1indices0indices1mutual0mutual1zeromatching_scores0matching_scores1valid0valid1rrs&& r(get_matches_from_scoresrVs||J1 !SbS#2#+  " "1 %D !SbS#2#+  " "1 %D||H||H||HNN1-hooFtLH||HNN1-hooFtLH//!X66G//!X66G ;;?? D ??1 D{{7$7{{7,C,CAx,PRVW 94 5F }}Q1 1F{{6R0H{{68R0Hkk8./99!Q?GG UVXZ[Gkk#3"FGQQRSUVW__`jmn`nprsO O ##r+cpV^8dQhR\PR\R\R\P/#)rrheightwidthrKry)r&s"r(r)r)s05<<SU\\r+c\P!W!.VPVPR7R,pV^, pVP R4P ^, pWR,, VR,, pV#)aw Normalize keypoints locations based on image_shape Args: keypoints (`torch.Tensor` of shape `(batch_size, num_keypoints, 2)`): Keypoints locations in (x, y) format. height (`int`): Image height. width (`int`): Image width. Returns: Normalized keypoints locations of shape (`torch.Tensor` of shape `(batch_size, num_keypoints, 2)`). rrmNrR).Nr|).NN)r"tensorrrmr@rD)rrXrYsizeshiftscales&&& r(normalize_keypointsr`sf << 0@0@  XY] ^D 1HE HHRL  ! #E<00E/4JJI r+zV LightGlue model taking images as inputs and outputting the matching of them. caa]tRtRtoRtV3RlV3RlltV3RlRltRV3RlRlltV3R lR ltRR lt V3R lR lt V3RlRlt Rt V3RlRlt RV3RlRllt]]RV3RlRll44tRtVtV;t#)LightGlueForKeypointMatchingiaJ LightGlue is a model matching keypoints in images by leveraging detections from a keypoint detector such as SuperPoint. It is based on the SuperGlue architecture and is designed to be lightweight and efficient. It consists of : 1. Keypoint Encoder 2. A Graph Neural Network with self and cross attention layers 3. Matching Assignment layers The correspondence ids use -1 to indicate non-matching points. Philipp Lindenberger, Paul-Edouard Sarlin and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV 2023. https://huggingface.co/papers/2306.13643 c <V^8dQhRS[/#r9r)r&r's"r(r))LightGlueForKeypointMatching.__annotate__sr+c 8<\SV`V4\P!VP4VnVPP VnVPVnVPVn VPVn VPVn VPVn VPVP8wd4\P!VPVPRR7VnM\P"!4Vn\%V4Vn\P(!\+VP4Uu.uFp\-WR7NK up4Vn\P(!\+VP4Uu.uFp\1V4NK up4Vn\P(!\+VP^, 4Uu.uFp\5V4NK up4VnVP94R#uupiuupiuupi)Tr=)rN)r?r@r from_configkeypoint_detector_configkeypoint_detectordescriptor_decoder_dim keypoint_detector_descriptor_dimrBnum_hidden_layers num_layersfilter_thresholddepth_confidencewidth_confidencerrAinput_projectionIdentityr7positional_encoder ModuleListrangertransformer_layersrmatch_assignment_layersr&token_confidence post_init)rFr:irHrGs&& r(r@%LightGlueForKeypointMatching.__init__s  !>!J!J6KjKj!k060O0O0f0f-$33 22 & 7 7 & 7 7 & 7 7   $"G"G G$&IId.S.SUYUhUhos$tD !$&KKMD !"s&& r(_get_confidence_threshold6LightGlueForKeypointMatching._get_confidence_thresholds;#tk'9DOO'K LLL wwy!Q''r+c <V^8dQhRS[PRS[PRS[R,RS[S[PS[S[PS[P3,3,/#)rrrrJNrKrL)r&r's"r(r)rds[?? <<?49LL?X\_cXc? u||U5<<#=>> ??r+cVP4P4pVPV4pVPW#R7pWE3#)rJ)r.rrprr)rFrrrJprojected_descriptorskeypoint_encoding_outputs&&&& r(_keypoint_processing1LightGlueForKeypointMatching._keypoint_processingsI"((*557 $ 5 5k B#'#:#:9#:#p $>>r+c <V^8dQhRS[PRS[RS[PRS[PRS[P/#)rkeypoint_confidencesr|r num_pointsrKry)r&r's"r(r)rd!sG##$)LL#?B#JO,,#didpdp# #r+c VPwrVW P^, 8dVPV^8H^4pVPV^,R4pVP V4pRW8P 4P ^R7V, , pWP8p V #\P!V\PR7p V #)zRevaluate whether we should stop inference based on the confidence of the keypointsg?rPrlrR) r}rlrrrrnsumrnr"onesrN) rFrr|rrrrHr>ratio_confidentearly_stopped_pairss &&&&& r(_get_early_stopped_image_pairs;LightGlueForKeypointMatching._get_early_stopped_image_pairs!s 1, ,$8#C#CDAIq#Q #7#?#? aQS#T 66{CI!%9%E$L$L$N$R$RWX$R$Y\f$ffO"14I4I"I  #"#(**Zuzz"J ""r+cVeW,pW$,pVPV,!W4p\WPP4wrgWg3#r)rvrVrm)rFrrr| early_stopsrrrs&&&&& r(_get_keypoint_matching3LightGlueForKeypointMatching._get_keypoint_matching4sI  "%2K$D--k:;M#:6CXCX#Y ''r+cn<V^8dQhRS[PRS[PRS[RS[P/#)r confidencesrr|rKry)r&r's"r(r)rd<s9U\\5<<^afkfrfrr+c nV^VP, 8pVeWAVPV48*,pV#)z#mask points which should be removed)ror)rFrrr|keeps&&&& r(_get_pruning_mask.LightGlueForKeypointMatching._get_pruning_mask<s:T2223  " 4#A#A+#NN ND r+c<V^8dQhRS[PRS[PRS[PRS[PRS[PRS[PRS[/#)rrrrrA prune_outputrr|ry)r&r's"r(r)rdCsr#_#_\\#_<<#_ll #_  #_ ll #_$ll#_#_r+c aVPwpp VPV,PV4p VPWjV4oSP V^8H\ P !R44oV3RlW^,V^,SV34wrrp\V4F!pVVVV,3;;,^, uu&K# RWW34wrrW3p\VRRR7pV VWV3#)zv For a given layer, prune keypoints based on the confidence of the keypoints and the matchability of the descriptors. Fc3|<"TF+p\VS4UUu.uF wr#W#,NK uppxK- R#uuppi5ir)zip).0r\trpruned_keypoints_masks& r( JLightGlueForKeypointMatching._do_layer_keypoint_pruning..Ws>c c%(0E$F G$FQWW$F Gc Hs <6<c3<"TFp\VRR7xK R#5i)T) batch_firstNr)r pruned_tensors& r(rr_s"S !j  D 9 9!jsTr padding_valuerR) r}rvr#rrr"r\rtr)rFrrrrArrr|rrHdescriptors_matchabilitypruned_descriptorspruned_keypoints_0pruned_keypoints_1 pruned_maskpruned_indicesrypruned_keypointsrs&&&&&&&& @r(_do_layer_keypoint_pruning7LightGlueForKeypointMatching._do_layer_keypoint_pruningCs ',, Aq#'#?#? #L#]#]^i#j $ 6 67Kgr s 5 A A$!)U\\Z_M` ac &! ilDY[bcc _0BQ_z"A N1-- .! 3 .#S "4J\!jS O0B/C%n$VXY!#3^R^^^r+ca\P!S4o\P!SP^,4pSV,pSV,oRWB34wrBRWS34wrSV3RlVVVV34wrEr#W#WE3#)rc3>"TFp\VRRR7xK R#5i)TrNrRrrr\s& r(rMLightGlueForKeypointMatching._concat_early_stopped_outputs..us#3 C T D DCc3>"TFp\VR^R7xK R#5i)TrNrrs& r(rrys#> N T C CNrc36<"TFpVS,xK R#5irr,)rr\early_stops_indicess& r(rr}s&g  & ' 's)r"rdrBr})rFrfinal_pruned_keypoints_indices!final_pruned_keypoints_iterationsrrids order_indicess&f&&&& r(_concat_early_stopped_outputs:LightGlueForKeypointMatching._concat_early_stopped_outputshs$kk*=>ll.44Q78+C0 1-@3 "C3 /> *N> :g .1 g c"@.RYjjr+c <V^8dQhRS[PRS[PRS[PRS[PRS[S[PS[P3,/#)rrArrrrK)r"rMr%)r&r's"r(r)rds_#*#*#*#* #* || #* u||U\\) * #*r+c raVPwopV3RlWV34wrpVR,pVR,pVR,pVR,p VR,p VR,p \P!S^,^V3RVPVPR7p \P !S^,^V3VPVPR7p \ S^,4Fp\P!W,R8HRW~,P^W,P^R744W^Wn,3&\P!W,R8HRWn,P^W,P^R744W^W~,3&W,W^Wn,3&W,W^W~,3&K W3#)c3V<"TFqPS^,^R4xK R#5i)rNrR)r)rr\rs& r(rJLightGlueForKeypointMatching._do_final_keypoint_pruning..s'- AdvNN:?Ar 2 2Ads&)r[)rrrrR) r}r"rrrmzerosrtrGrCclamp)rFrArrrrHrMrNrKrLrRrS_matches_matching_scoresryrs&&&&& @r(_do_final_keypoint_pruning7LightGlueForKeypointMatching._do_final_keypoint_prunings A- BITcAd- )/4=4=4=4=*40*40::zQ=A2gnndkdqdqr ;; 1_a /oNcNc zQ'A*/++ r!2x{'9'9!X[=N=NST=N=U'V+H8;& '+0++ r!2x{'9'9!X[=N=NST=N=U'V+H8;& '3C2E 8;. /2B2E 8;. /())r+c<V^8dQhRS[PRS[PRS[RS[RS[PR,RS[R,RS[R,R S[S[PS[PS[PS[S[3,/#) rrrrXrYrNrrJrK)r"rMrzrNr%)r&r's"r(r)rdsq q <<q \\q  q  q llT! q  $;q #Tkq  u||U\\5<<E Fq r+c ( a'V'dR MRpV'dR MRp VP^,^8XdUVPRR p VPV R \PR7VP V 4VP V 4VV 3#VP p VPwrr\P !VPV R 4^R7pVPV ^,V^4pVeVPV ^,V4MRpVPV ^,WP4p\P!V ^,V R7p\WV4pVPW!VR7wppV^,pVP^8pVP^8p.p.p.p.p.p\P!^WR7PV ^,R 4p\P!V4p\!VP"4EFpVe"\%VP&VR,VRR7pM8\P(!WP+4R,3VP R7pVP,V,!VVVVVR7pVwpppV'd VV,pV'd V V,p V'EdVVP"^, 8d/VP.V,!V4p VP1V VW_R7p!M&\P(!V \P2R7p!\P4!V!4'EdxV!P7^4o'VS',p"VP9W%VS'R 7wp#p$VP;\=V"44VP;\=V#44VP;\=V$44V'dCVP;\=VS',44VP;\=VS',44VV!(,p\>;QJd).V'3R lW!^,V^,VV34FNK 5M"!V'3R lW!^,V^,VV344wpp%p&ppV%V&3pV'd=\>;QJd.V'3R lVVX 34FNK 5M!V'3R lVVX 344wppp \P@!V!4'dM,V'gEKVPCVVVVVX V4wr!pppEK V'd;V'd3VPEVVVVV4wppppVPGVVVV4wppMMVP9W%VP"^, 4wpp\P!V4VP",pVPV ^V4pVVVVV 3#)NrlrPrrF)r: inputs_embedsrallow_is_bidirectional_skip)rrJr)r)rc38<"TFpVS(,xK R#5irr,rr\rs& r(rALightGlueForKeypointMatching._match_image_pair..s#V&dF |,,&dc38<"TFpVS(,xK R#5irr,rs& r(rr#s'l+#K<00+rr,rR)r|:rrbNr|rc)$r}rr"rz new_zerosrrrrjrBr`rrnror~ ones_likertrlr r:rr]rurwrrNanyrSrextendlistr%allrrr)(rFrrrXrYrrrJrrr}rrrHinitial_num_keypointsnum_points_per_pair image_indicesr do_early_stopdo_keypoint_pruningrrrrrpruned_keypoints_indicespruned_keypoints_iterationsr|extended_attention_mask layer_outputr attentionrrearly_stopped_image_indicesearly_stopped_matchesearly_stopped_matching_scores keypoints_0 keypoint_1rs(&&&&&&&& @r(_match_image_pair.LightGlueForKeypointMatching._match_image_pairs#7BD0d ??1  "OOCR(E""5"EII">##E*##E*!  !!2;/// ,#ii Z(D!L%%j1n6KQO FJFVt||JN,AB\`!))*q.:OQvQvw  Z!^FC ' 5A 040I0I 9M1J1 - --Q/ --1 #33a7 )+&,.)#(<<3H#X#_#_`jmn`npr#s &+oo6N&O# 1K*C;;"-i"8#'05 +'+0**jBRBRBTUWBX5Ybkbrbr*s'22;?6%9"3 L5A 1K #$5 $E! !/)!;}1!44+/+@+@+Mk+Z(+/*M*M,k4+N+' +0**Zuzz*R'99011#6"G"G"JK2? 2L/KOKfKf#;KLgLH)+H(..t4O/PQNN4(=#>?#**40M+NO*6==dC[\gCh>ij9@@FabmFnAop+>?R>R*S'PUPUV'2aL)A,PTVc&dVPUPUV'2aL)A,PTVc&dVQMKj$ "-j 9I*fkfkl!9 ; 4+lfkfkl!9 ; 4+lgc02MOc99011""33#!03,#d (@$Hc]2t 022'25#  h *,MwXg(,'F'F.% ( $G_(,'B'B;VZVeVehiVi'j $G_050PSWSbSb0b -,M,U,U 0- )   -     r+c <V^8dQhRS[PRS[PR,RS[R,RS[R,RS[S[,/#)rr3labelsNrrJrK)r"r# LongTensorrNr%r)r&r's"r(r)rdbs\4 4 ''4   4'4  $; 4 #Tk 4  0 04 r+c RpVe \R4hVeTMVPPpVeTMVPPpVP^8wgVP ^4^8wd \R4hVP wrxrp VPV^,WV 4pVPV4p V R,wrrV PV^R^4PV4p VPV^RVP4PV4pVPV^R4pV P4pVR,V ,VR&VR,V ,VR&VPVVV V VVVR7wppppp\VVVV VVVVR7#) Nz9LightGlue is not trainable, no labels should be provided.zOInput must be a 5D tensor of shape (batch_size, 2, num_channels, height, width):NN)rrrJ)rrrrrrr r!rR)r|r|r|r)r|r|r|rb) ValueErrorr:rrJndimr]r}rrhrprjclonerr)rFr3rrrJrrrrHchannelsrXrYkeypoint_detectionsrrrabsolute_keypointsrrrr r!s&&&&&, r(r[$LightGlueForKeypointMatching.forward`s  XY Y1B1N-TXT_T_TqTq$8$D $++JjJj     !\%6%6q%9Q%>no o1=1C1C. x#++JNHeT "44\B*=b*A' k%%j!R;>>|L !))*aT=b=bcffgst ||J2.&__.);J)G%)O:&);J)G&)P:&EIE[E[    /!5F\F B% /+'!  r+) rnrBrmrprhrjrvrlrrrwruror]r)NNN)r-r.r/r0r1r@rrrrrrrrrrrr[r3r4r_r`s@@r(rbrbs <(( ??##&(#_#_Jk@#*#*Jq q f4 4 4 4 r+rb)rb)r)0collections.abcr dataclassesrnumpyr~r"rtorch.nn.utils.rnnr activationsr masking_utilsr modeling_flash_attention_utilsr modeling_utilsr r processing_utilsr utilsrrrrauto.modeling_autorconfiguration_lightgluerrrr7rjrvrrrrrrrr&r1rVr`rb__all__r,r+r(rsB(%! +!6BF&VV>4  7k 7  7F"<8 U%2D)D)N299"P> P>f &BII&R BII    $@, o #;o  o d &'E Fr+