R@S5r!Sr""SS\ R:5r#S2S\RHS\%S\&S\RH4Sjjr'"SS\ R:5r("SS\ R:5r)"SS\ R:5r*"SS \ R:5r+S!r,S"r-"S#S$\ 5r."S%S&\ R:5r/S'\ R:SS4S(jr0\"S)S*\55r1\"S+S,\155r2\"S-S.9"S/S0\1\55r3/S1Qr4g)3zPyTorch ViTDet backbone.N)OptionalUnion)nn)ACT2FN)GradientCheckpointingLayer)BackboneOutputBaseModelOutput)PreTrainedModel)auto_docstringlogging) BackboneMixin) VitDetConfigcl^\rSrSrSrU4SjrSrS\RS\R4Sjr Sr U=r $) VitDetEmbeddings%z This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial `hidden_states` (patch embeddings) to be consumed by a Transformer. cx>[TU]5 URURp2URUR pT[ U[RR5(aUOX"4n[ U[RR5(aUOX34nUSUS-USUS--nX l X0lX@lX`l UR(a@US-n[R"[R "SXqR 55UlOSUl[R$"XEX3S9Ulg)Nrr) kernel_sizestride)super__init__pretrain_image_size patch_size num_channels hidden_size isinstance collectionsabcIterable image_size num_patches use_absolute_position_embeddingsr Parametertorchzerosposition_embeddingsConv2d projection) selfconfigr!rrrr" num_positions __class__s b/var/www/html/shao/venv/lib/python3.13/site-packages/transformers/models/vitdet/modeling_vitdet.pyrVitDetEmbeddings.__init__+s !'!;!;V=N=NJ$*$7$79K9Kk#-j+//:R:R#S#SZZdYq #-j+//:R:R#S#SZZdYq !!} 1 5*Q-:VW=:XY $$(&  2 2'!OM')||EKK=RdRd4e'fD $'+D $))L:icU(a USS2SS24nURSn[[R"U55nXf-U:wa [ S5e[ R R5(d Xc:wdXd:waX[RRURSXfS5RSSSS5X44SS S 9nURSSSS5$URSX4S5$) a? Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token dimension for the original embeddings. Args: abs_pos_embeddings (`torch.Tensor`): Absolute positional embeddings with (1, num_position, num_channels). has_cls_token (`bool`): If true, has 1 embedding in abs_pos_embeddings for cls token. height (`int`): Height of input image tokens. width (`int`): Width of input image tokens. Returns: Absolute positional embeddings after processing with shape (1, height, width, num_channels) Nrz5Absolute position embeddings must be a square number.rrbicubicF)sizemode align_corners) shapeintmathsqrt ValueErrorr%jit is_tracingr functional interpolatereshapepermute)r*abs_pos_embeddings has_cls_tokenheightwidth num_positionr5new_abs_pos_embeddingss r.get_absolute_positions'VitDetEmbeddings.get_absolute_positionsAs$ !3AqrE!: )//2 499\*+ ;, &TU U 99   ! !dn %']]%>%>"**1d"=EEaAqQ_# &?& "*11!Q1= =%--aC Cr0 pixel_valuesreturnczURSnX R:wa[SURSUS35eURU5nURbbUR SSSS5nX0R URSURSURS5-nUR SSSS5nU$) NrzoMake sure that the channel dimension of the pixel values match with the one set in the configuration. Expected z but got .rr3rT)r8rr<r)r'rBrI)r*rKr embeddingss r.forwardVitDetEmbeddings.forwardgs#))!, ,, ,!../yaI __\2  # # /#++Aq!Q7J#&A&A(($ 0@0@0CZEUEUVWEX'J$++Aq!Q7Jr0)r!rr"rr'r)) __name__ __module__ __qualname____firstlineno____doc__rrIr%TensorrP__static_attributes__ __classcell__r-s@r.rr%s5 j,$DLELLU\\r0rcH[S[X5-S- 5nURSU:waq[RR UR SURSS5RSSS5USS9nUR SU5RSS5nOUn[R"U5SS2S4[X- S5-n[R"U5SSS24[X- S5-nXV- US- [X- S5--nXGR5$) aq Get relative positional embeddings according to the relative positions of query and key sizes. Args: q_size (`int`): Size of query q. k_size (`int`): Size of key k. rel_pos (`torch.Tensor`): Relative position embeddings (num_embeddings, num_channels). Returns: Extracted positional embeddings according to relative positions. r3rrr2linear)r5r6N?) r9maxr8rr?r@rArBr%arangelong)q_sizek_sizerel_pos max_rel_distrel_pos_resizedq_coordsk_coordsrelative_coordss r. get_rel_posri}s q3v..23L}}Q<'--33 OOAw}}Q/ 4 < bhwkzbhwc,wkc->bhwkN)rir8rAr%einsumview)attnqueries rel_pos_h rel_pos_wrarbqueries_height queries_width keys_height keys_widthrelative_heightrelative_width batch_size_dimr_qrelative_weights r.!add_decomposed_relative_positionsr|s,%+!N$K!.yIO IFN J3 //*m ICll#3SJOll#3SIO *m*U !Q1d* + , !Q4* + , d: 5{7OP  Kr0c:^\rSrSrSrSU4SjjrSSjrSrU=r$)VitDetAttentionz=Multi-head Attention block with relative position embeddings.c(>[TU]5 URnURnX@lX4-nUS-Ul[ R"X3S-URS9Ul [ R"X35Ul URUl UR(as[ R"[R"SUS-S- U55Ul[ R"[R"SUS-S- U55Ulgg)z Args: config (`VitDetConfig`): Model configuration. input_size (`tuple[int]`, *optional*): Input resolution, only required in case relative position embeddings are added. grbiasr3rrN)rrrnum_attention_heads num_headsscalerLinearqkv_biasqkvproj use_relative_position_embeddingsr$r%r&rorp)r*r+ input_sizeryrhead_dimr-s r.rVitDetAttention.__init__s   .. "#t^ 99S'@IIc' 060W0W-  0 0\\%++a*Q-6G!6KX*VWDN\\%++a*Q-6G!6KX*VWDN 1r0c URup4pVURU5RX4U-SURS5R SSSSS5nURSX0R-XE-S5R S5upn XR -U RSS5-n UR(a%[XURURXE4XE45n U RSS9n X-nURX0RXES5nUR SSSSS5nURX4US5nURU5nU(a<U RX0RU RSU RS5n X4n U $U4n U $) Nrr2r3rr)ry)r8rrArrBunbindr transposerr|rorpsoftmaxrlr)r* hidden_stateoutput_attentionsrwrErFrxrrnkeysvaluesattention_scoresattention_probsoutputss r.rPVitDetAttention.forwards'3'9'9$ Ehh|$,,Z%DNN\^_gghiklnoqrtuv # AzNN/JFN\^ _ f fgh iv#jj0DNN2r4JJ  0 0@ 4>>4>>F?]c\k  +22r2:&/ #((^^VTVW #++Aq!Q: #++JrJ yy. -55NNO,A,A",EG\G\]_G`O$5G$oGr0)rrrrorprrN)F rRrSrTrUrVrrPrXrYrZs@r.r~r~sGX4r0r~input drop_probtrainingrLcUS:XdU(dU$SU- nURS4SURS- --nU[R"X@RUR S9-nUR 5 URU5U-nU$)a* Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks, however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper... See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the argument. rr)r)dtypedevice)r8ndimr%randrrfloor_div)rrr keep_probr8 random_tensoroutputs r. drop_pathr sCx I I [[^  Q 7 7E 5 ELL YYM YYy !M 1F Mr0c^\rSrSrSrS S\\SS4U4SjjjrS\RS\R4Sjr S\ 4S jr S r U=r$) VitDetDropPathizXDrop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).NrrLc.>[TU]5 Xlgr)rrr)r*rr-s r.rVitDetDropPath.__init__"s "r0 hidden_statescB[XRUR5$r)rrr)r*rs r.rPVitDetDropPath.forward&s FFr0c SUR3$)Nzp=rr*s r. extra_reprVitDetDropPath.extra_repr)sDNN#$$r0rr)rRrSrTrUrVrfloatrr%rWrPstrrrXrYrZs@r.rrsQb#(5/#T##GU\\GellG%C%%r0rc6^\rSrSrSrSU4SjjrSrSrU=r$)VitDetLayerNormi-a< A LayerNorm variant, popularized by Transformers, that performs point-wise mean and variance normalization over the channel dimension for inputs that have shape (batch_size, channels, height, width). https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119 c>[TU]5 [R"[R "U55Ul[R"[R"U55UlX l U4Ul gr) rrrr$r%onesweightr&repsnormalized_shape)r*rrr-s r.rVitDetLayerNorm.__init__4sR ll5::.>#?@ LL-=!>? !1 3r0cURSSS9nX- RS5RSSS9nX- [R"X0R-5- nUR SS2SS4U-UR SS2SS4-nU$)NrT)keepdimr3)meanpowr%r;rrr)r*xuss r.rPVitDetLayerNorm.forward;s FF1dF # UKKN  4  0 UejjXX. . KK4 & *TYYq$}-E Er0)rrrr)gư>rrZs@r.rr-s 4r0rc2^\rSrSrSrU4SjrSrSrU=r$)VitDetResBottleneckBlockiCz The standard bottleneck residual block without the last activation layer. It contains 3 conv layers with kernels 1x1, 3x3, 1x1. c>[TU]5 [R"X$SSS9Ul[ U5Ul[URUl [R"XDSSSS9Ul [ U5Ul [URUl [R"XCSSS9Ul [ U5Ulg)a" Args: config (`VitDetConfig`): Model configuration. in_channels (`int`): Number of input channels. out_channels (`int`): Number of output channels. bottleneck_channels (`int`): Number of output channels for the 3x3 "bottleneck" conv layers. rFrr)paddingrN)rrrr(conv1rnorm1r hidden_actact1conv2norm2act2conv3norm3)r*r+ in_channels out_channelsbottleneck_channelsr-s r.r!VitDetResBottleneckBlock.__init__Is YY{O $%89 6,,- YY2TU\ab $%89 6,,- YY2!%P $\2 r0cPUnUR5H nU"U5nM X-nU$r)children)r*routlayers r.rP VitDetResBottleneckBlock.forwardas.]]_E*C%g r0)rrrrrrrrrrZs@r.rrCs 30r0rcr^\rSrSrS\S\SS4U4SjjrS\RS\R4SjrS r U=r $) VitDetMlpij in_featureshidden_featuresrLNc>[TU]5 [R"X#5Ul[ UR Ul[R"X25Ul[R"UR5Ul gr) rrrrfc1rractfc2Dropout dropout_probdrop)r*r+rrr-s r.rVitDetMlp.__init__ksV 99[:&++,99_:JJv223 r0rcURU5nURU5nURU5nURU5nURU5nU$r)rrrr)r*rs r.rPVitDetMlp.forwardrsH HHQK HHQK IIaL HHQK IIaLr0)rrrr) rRrSrTrUr9rr%rWrPrXrYrZs@r.rrjs=4C4#4$4%,,r0rc HURup#pEXU-- U-nXU-- U-n[RRUSSSUSU45nX6-XG-pUR X(U-XU-X5nUR SSSSSS5R 5R SXU5n XU 44$)a Partition into non-overlapping windows with padding if needed. Args: hidden_state (`torch.Tensor`): Input tokens with [batch_size, height, width, num_channels]. window_size (`int`): Window size. Returns: `tuple(torch.FloatTensor)` comprising various elements: - windows: windows after partition with [batch_size * num_windows, window_size, window_size, num_channels]. - (padded_height, padded_width): padded height and width before partition rrrr3rr2)r8rr?padrlrB contiguous) r window_sizerwrErFr pad_height pad_width padded_height padded_widthwindowss r.window_partitionr|s/;.@.@+J 44 CJ{22kAI==$$\Aq!Y:3VWL"("5u7H<$$[0+{?Z\gL""1aAq!4??AFFr;eqrG L1 11r0c(UupEUupgURSXE-U-U--nURXU-XQ-XS5n U RSSSSSS5R5n U RXUS5n U SS2SU2SU2SS24R5n U $) a Window unpartition into original sequences and removing padding. Args: windows (`torch.Tensor`): Input tokens with [batch_size * num_windows, window_size, window_size, num_channels]. window_size (`int`): Window size. pad_height_width (`tuple[int]`): Padded height and width (padded_height, padded_width). height_width (`tuple[int]`): Original height and width before padding. Returns: hidden_state: unpartitioned sequences with [batch_size, height, width, num_channels]. rr2rrr3rrN)r8rlrBr) rrpad_height_width height_widthrrrErFrwrs r.window_unpartitionrs"#3M MFq!m&Bk&QU`&`aJ<<[0,2M{ikL ''1aAq9DDFL$$Z bQL 7F7FUFA 56AACL r0c ^\rSrSrSrSS\S\S\S\SS4 U4S jjjr SS \ RS \ \ RS \S\ \\ R\ R4\\ R44S jjrSrU=r$) VitDetLayerizCThis corresponds to the Block class in the original implementation.r+drop_path_rateruse_residual_blockrLNc>[T U]5 URnURn[ U[ [ 45(aUOXf4nURn[ U[ [ 45(aUOXw4nUSUS-USUS-4n[R"XQRS9Ul [XS:XaUOX34S9Ul US:a [U5O[R"5Ul[R"XQRS9Ul[%X['XQR(-5S9UlX0lX@lUR.(a[1UUUUS-S9Ulgg) Nrr)r)rr)r+rrr3)r+rrr)rrrr!rlisttuplerr LayerNormlayer_norm_epsrr~ attentionrIdentityrrrr9 mlp_ratiomlprrrresidual) r*r+rrrryr!rrr-s r.rVitDetLayer.__init__sF   && #-j4-#H#HZzNf && #-j4-#H#HZzNf  mz!}4jmzRS}6TU \\#+@+@A ( A-=zKC] }N^N^._ +M!% /"0" /q1 (,   a .}>N>NP`cikpbqrM!>>-#@@ %txx =@Y7Z([[ %--aAq9  " " MM-8M "W,r0)rrr rrr rr)rrF)NF)rRrSrTrUrVrrr9boolrr%rWrrrrPrXrYrZs@r.rrsMqv!"!49!LO!im! !!L-1"' (||(ELL)( ( uU\\5<</0% 2EE F ((r0rc^\rSrSrS\SS4U4SjjrS S\RS\\RS\ S \ S \ S\ \ \ 44 S jjr S rU=r$) VitDetEncoderi r+rLNc >[TU]5 XlURn[R "SUR USS9Vs/sHo3R5PM nn/n[U5HGnUR[UXFXaR;a UROSXaR;S95 MI [R"U5UlSUlgs snf)Nrcpu)r)rrrF)rrr+num_hidden_layersr%linspaceritemrangeappendrwindow_block_indicesrresidual_block_indicesr ModuleListrgradient_checkpointing)r*r+depthrrlayersir-s r.rVitDetEncoder.__init__ s  ((-2NN1f>S>SUZch,ij,iq&&(,ijuA MM#1#467;V;V6V 2 2\]'(,I,I'I  ]]6* &+#ksCrr routput_hidden_states return_dictc6U(aSOSnU(aSOSn[UR5H9upU(aXa4-nUbX(OSn U "XU5n U SnU(dM1X{S4-nM; U(aXa4-nU(d[SXU455$[UUUS9$)Nrrc3,# UH ocMUv M g7frr').0vs r. (VitDetEncoder.forward..?sm$[q$[s last_hidden_stater attentions) enumeraterrr ) r*rr rr$r%all_hidden_statesall_self_attentionsr" layer_modulelayer_head_mask layer_outputss r.rPVitDetEncoder.forward#s#7BD$5b4(4OA#$58H$H!.7.CilO(IZ[M)!,M  &91=MR RR5 UR>RRR5 gg)zInitialize the weightsr)rstdNr]) rrrr(r< trunc_normal_rdatator%float32r+initializer_rangerrzero_rfill_rr'r~rrorprrrrr?rrr)r*r7rs r. _init_weights#VitDetPreTrainedModel._init_weights]s fryy"))4 5 5"$!6!6 ""%%emm43DKKDaDa"7"b$$% MM {{&   &&('  - - KK   " " $ MM   $ $S )  0 1 1.0gg.C.C**//225==AKK11/D/b++112  & & +  0 0T[[5a5a$&GG$9$9  %%((7KK11%:%F   ! %'GG$9$9  %%((7KK11%:%F   !  8 9 9 ,, fllC 'D ,, 5 !!'', %%'6 LL   $ $ * * , LL   " " ( ( *:r0r')rRrSrTrUr__annotations__base_model_prefixmain_input_namesupports_gradient_checkpointing_no_split_modulesrrrr(rrLrXr'r0r.rArAUsK  $O&*#)+E"))RYY *L$M)+RV)+r0rAc^\rSrSrS\4U4SjjrS\4SjrS\\ \ \ 4SS4Sjr \ SS \ \RS \ \RS \ \S \ \S \ \S\\\44 Sjj5rSrU=r$) VitDetModelir+c>[TU]U5 Xl[U5Ul[ U5UlUR5 gr)rrr+rrOrencoder post_init)r*r+r-s r.rVitDetModel.__init__s9   *62$V,  r0rLc.URR$rrOr)rs r.get_input_embeddings VitDetModel.get_input_embeddings)))r0heads_to_pruneNcUR5H7up#URRURR U5 M9 g)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel N)itemsrVrr prune_heads)r*r^rheadss r. _prune_headsVitDetModel._prune_headss< +002LE LL  u % / / ; ;E B3r0rKr rr$r%cUbUOURRnUbUOURRnUbUOURRnUc [ S5eUR X RR 5nURU5nURUUUUUS9nUSnU(d U4USS-$[UURURS9$)aw Examples: ```python >>> from transformers import VitDetConfig, VitDetModel >>> import torch >>> config = VitDetConfig() >>> model = VitDetModel(config) >>> pixel_values = torch.randn(1, 3, 224, 224) >>> with torch.no_grad(): ... outputs = model(pixel_values) >>> last_hidden_states = outputs.last_hidden_state >>> list(last_hidden_states.shape) [1, 768, 14, 14] ```Nz You have to specify pixel_values)r rr$r%rrr-) r+rr$use_return_dictr< get_head_maskrrOrVr rr/) r*rKr rr$r%embedding_outputencoder_outputssequence_outputs r.rPVitDetModel.forwards82C1N-TXT_T_TqTq$8$D $++JjJj &1% = 'tn = d^ =  uo% &= = r0rTzF ViTDet backbone, to be used with frameworks like Mask R-CNN. ) custom_introc ^\rSrSrU4SjrS\4Sjr\S S\RS\ \ S\ \ S\ \ S\ 4 S jj5r S rU=r$) VitDetBackboneic>[TU]U5 [TU] U5 [U5Ul[ U5Ul[URS-5Vs/sHo!RPM snUl UR5 gs snf)Nr) rr_init_backbonerrOrrVrrr num_featuresrW)r*r+rxr-s r.rVitDetBackbone.__init__su   v&*62$V, 9>v?W?WZ[?[9\]9\A//9\] ^sBrLc.URR$rrZrs r.r[#VitDetBackbone.get_input_embeddingsr]r0rKr$rr%c UbUOURRnUbUOURRnUbUOURRnUR U5nUR USUUS9nU(a UR OUSnSn[URU5HupXR;dMX4- nM U(dU(a U4USS-n U $U4USS-n U $[UU(a UR OSURS9$)ap Examples: ```python >>> from transformers import VitDetConfig, VitDetBackbone >>> import torch >>> config = VitDetConfig() >>> model = VitDetBackbone(config) >>> pixel_values = torch.randn(1, 3, 224, 224) >>> with torch.no_grad(): ... outputs = model(pixel_values) >>> feature_maps = outputs.feature_maps >>> list(feature_maps[-1].shape) [1, 768, 14, 14] ```NT)r$rr%rr'r3) feature_mapsrr/) r+rfr$rrOrVrzip stage_names out_featuresr r/) r*rKr$rr%rhrrrwstagerrs r.rPVitDetBackbone.forwards+6&1% ; d^ ;  ; ; r0ro)rTrAro)rF)5rVcollections.abcrr:typingrrr%torch.utils.checkpointr activationsrmodeling_layersrmodeling_outputsr r modeling_utilsr utilsr r utils.backbone_utilsrconfiguration_vitdetr get_loggerrRloggerModulerr=script_if_tracingrir|r~rWrrrrrrrrrrrr?rArTro__all__r'r0r.rs " !9?-,1.   H %UryyUp!3!3H&R;bii;~U\\eTV[VbVb* %RYY %bii,$ryy$N $2@>N,Nb8 BII8 v *RYY *4 *0+O0+0+fT 'T T n K *MK  K \ Er0