RЦi9R XSrSSKJr SSKJrJrJrJrJrJ r SSK r SSK J r SSK JrJr SSKJrJrJrJrJrJrJrJr SSKJr SS KJr SS KJr SS KJ r J!r! S /r""S S\ RF5r$"SS\ RF5r%S\&4Sjr'"SS\ RF5r("SS\ RR5r*"SS \ RF5r+S+Sjr,Sr-S,Sjr.S-Sjr/\!"\/"SS9\/"SS9\/"S S9\/"S!S9\/"S"S#SS$S%9S&.5r0\ S,S'j5r1\ S,S(j5r2\ S,S)j5r3\ S,S*j5r4g).z? RDNet Copyright (c) 2024-present NAVER Cloud Corp. Apache-2.0 )partial)ListOptionalTupleUnionCallableTypeNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)DropPathcalculate_drop_path_ratesNormMlpClassifierHeadClassifierHeadEffectiveSEModulemake_divisible get_act_layerget_norm_layer)build_model_with_cfg)feature_take_indices) named_apply)register_modelgenerate_default_cfgsRDNetc ^\rSrSrS S\S\S\S\\RS\\R4 U4SjjjrSr S r U=r $) Blockin_chs inter_chsout_chs norm_layer act_layerc >XgS.n[T U]5 [R"[R"X4USSSS.UD6U"U40UD6[R"X4SSSS.UD6U"5[R"X#4SSSS.UD65UlgNdevicedtyper)groups kernel_sizestridepaddingrr,r-r.)super__init__nn SequentialConv2dlayers selfrr r!r"r#r'r(dd __class__s P/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/timm/models/rdnet.pyr1Block.__init__s / mm IIf ^V1VW ^[] ^ v $ $ IIf RQq! Rr R K IIi Sa1 SPR S   c$URU5$Nr5r7xs r:forward Block.forward-{{1~r<r?NN __name__ __module__ __qualname____firstlineno__intr r2Moduler1rB__static_attributes__ __classcell__r9s@r:rrs_     RYY  BII   (r<rc ^\rSrSrS S\S\S\S\\RS\\R4 U4SjjjrSr S r U=r $) BlockESE1rr r!r"r#c .>XgS.n[T U]5 [R"[R"X4USSSS.UD6U"U40UD6[R"X4SSSS.UD6U"5[R"X#4SSSS.UD6[ U40UD65Ulgr%)r0r1r2r3r4rr5r6s r:r1BlockESE.__init__2s / mm IIf ^V1VW ^[] ^ v $ $ IIf RQq! Rr R K IIi Sa1 SPR S g , ,   r<c$URU5$r>r?r@s r:rBBlockESE.forwardGrDr<r?rErFrOs@r:rQrQ1s_     RYY  BII   *r<rQblockcUR5R5nUS:Xa[$US:Xa[$SUS35e)NrWblockesezUnknown block type (z).)lowerstriprrQ)rWs r:_get_block_typer\KsD KKM   !E  * 6,UG266ur<c^\rSrSrSSSSSSSSS\R \R SS4 S \S \S \S \S \S\S\S\ S\S\ \RS\ \R4U4Sjjjr S\ \RS\R4SjrSrU=r$) DenseBlockU@@rrư>Nnum_input_features growth_ratebottleneck_width_ratiodrop_path_rate drop_raterand_gather_step_prob block_idx block_type ls_init_valuer"r#c p>XS.n[TU]5 XPlX@lX`lXplX lU S:a.[R"U [R"U40UD6-5OSUl [U5n[X-S- 5S-n[U5Ul[U5"SUUUU U S.UD6Ulg)Nr&r)rr r!r"r#)r0r1rhrgrirjrer2 ParametertorchonesgammarKr drop_pathr\r5)r7rdrerfrgrhrirjrkrlr"r#r'r(r8r r9s r:r1DenseBlock.__init__Vs  / ",%:""&TadeTeR\\-%**[2OB2O"OPko +& *CaGH1L !.1%j1 %!     r<rAreturnc[R"US5nURU5nURb-UR URR SSSS55nUR U5nU$)Nr)rqcatr5rsmulreshapertr@s r:rBDenseBlock.forward}s^ IIaO KKN :: !djj((B156A NN1 r<)rjrtrgrhrsrer5ri)rGrHrIrJr2 LayerNormGELUrKfloatstrr rLr1rrqTensorrBrMrNrOs@r:r^r^Us')!,/$'"+.%#'*,,,)+% #% % %* % " %  % $)% % % !% RYY% BII% % Nell+ r<r^c ^\rSrSrS S\S\S\\S\4U4SjjjrS\RS\R4S jr S r U=r $) DenseStage num_blockrddrop_path_ratesrec >XVS.n[T U]5 [U5H0n [SUUX9U S.UDUD6n X$- nUR SU 3U 5 M2 X lg)Nr&)rdrergrj dense_blockro)r0r1ranger^ add_modulenum_out_features) r7rrdrrer'r(kwargsr8ilayerr9s r:r1DenseStage.__init__s / y!A#5'.1     E  -  OOk!-u 5"!3r< init_featurervczU/nUHnU"U5nURU5 M [R"US5$)Nr)appendrqry)r7rfeaturesmodule new_features r:rBDenseStage.forwards< >F *K OOK (yy1%%r<)rrE) rGrHrIrJrKrrr1rqrrBrMrNrOs@r:rrs_33!$3"%[ 3  332&ELL&U\\&&r<rc+ ^\rSrSrS,S\S\S\S\\\\\4S\\\\\4S\\\\\4S\\\ \\ 4S \ S \ S \ S \S \S\S\ S\ S\ S\\\ 4S\S\ \ S\ S\ 4*U4Sjjjr \RR S-Sj5r\RR S.Sj5r\RR S\R(4Sj5rS/S\S\ \4SjjrS0S\R.S\ \\\\4S\ S \ S!\S"\ S\\\R.\\R.\\R.444S#jjrS1S\\\\4S$\ S%\ 4S&jjrS'rS-S(\ 4S)jjrS*rS+rU=r$)2rin_chans num_classes global_pool growth_ratesnum_blocks_listrkis_downsample_blockrftransition_compression_ratiorl stem_type patch_sizenum_init_featureshead_init_scalehead_norm_first conv_biasr#r"norm_epsrhrgc>[T%U]5 UUS.n[U5[U5s=:Xa[U5:Xde e[U5n[ U5nUb [ UUS9nX lXlUUlU S;deU S:Xa>[R"[R"X4XUS.UD6U"U 40UD65Ul U nOrSU ;a[U S-5OU n[R"[R"UU4S SS US .UD6[R"UU 4S SS US .UD6U"U 40UD65Ul S n/Ul[U5UlUnU n[!UUS S9n/n[#UR5GH8n/n US:wap[%UU -S- 5S-n!S =n"n#UU(a US-nS=n"n#U R'U"U40UD65 U R'[R"UU!4U"U#SS.UD65 U!n[)SUUUUUUUUUU UUUUS. UD6n$U R'U$5 UUUUU-- nUS -UR:Xd US -UR:wa3UUS -(a&U=R[+UUSU3UUS9/- slUR'[R"U 65 GM; [R"U6UlU=UlUlU(aBU"UR.40UD6Ul[5UR.U4UURS.UD6UlOD[R8"5Ul[;UR.U4UURUS.UD6Ul[=[ [>US9U5 g)aY Args: in_chans: Number of input image channels. num_classes: Number of classes for classification head. global_pool: Global pooling type. growth_rates: Growth rate at each stage. num_blocks_list: Number of blocks at each stage. is_downsample_block: Whether to downsample at each stage. bottleneck_width_ratio: Bottleneck width ratio (similar to mlp expansion ratio). transition_compression_ratio: Channel compression ratio of transition layers. ls_init_value: Init value for Layer Scale, disabled if None. stem_type: Type of stem. patch_size: Stem patch size for patch stem. num_init_features: Number of features of stem. head_init_scale: Init scaling value for classifier weights and biases. head_norm_first: Apply normalization before global pool + head. conv_bias: Use bias layers w/ all convolutions. act_layer: Activation layer type. norm_layer: Normalization layer type. norm_eps: Small value to avoid division by zero in normalization. drop_rate: Head pre-classifier dropout rate. drop_path_rate: Stochastic depth drop rate. r&N)eps)patchoverlapoverlap_tieredr)r,r-biastieredr*r)r,r-r.rT) stagewiserrnr/) rrdrerfrhrrlrkr"r#z dense_stages.)num_chs reductionrre) pool_typerh)rrhr")rro) r0r1lenrrrrrrhr2r3r4stemr feature_info num_stagesrrrKrrdict dense_stages num_featureshead_hidden_sizenorm_prerheadIdentityrr _init_weights)&r7rrrrrrkrrfrrlrrrrrrr#r"rrhrgr'r(r8 stem_stridemid_chs curr_striderdp_ratesrrdense_stage_layerscompressed_num_featuresk_sizer-stager9s& r:r1RDNet.__init__sb  /< C$8TC@S)rgrad_checkpointing)r7enabless r:set_grad_checkpointingRDNet.set_grad_checkpointingSs""A#) #r<rvc.URR$r>)rfc)r7s r:get_classifierRDNet.get_classifierXsyy||r<cFXlURRX5 gr>)rrreset)r7rrs r:reset_classifierRDNet.reset_classifier\s&  1r<rAindicesnorm stop_early output_fmtintermediates_onlycUS;dS5e/nURVs/sH#n[USRS5S5PM% n n[[ U 5U5upU V s/sHoU PM n n Xn UR U5n[ UR 5S- n [RR5(dU(d UR nOUR SU S-n[U5HFunnU"U5nX;dMU(aX:XaURU5nOUnURU5 MH U(aU$WU :XaURU5nX4$s snfs sn f)aForward features that returns intermediates. Args: x: Input image tensor indices: Take last n blocks if int, all if None, select matching indices if sequence norm: Apply norm layer to compatible intermediates stop_early: Stop iterating over blocks when last desired intermediate hit output_fmt: Shape of intermediate feature outputs intermediates_only: Only return intermediate features )NCHWzOutput shape must be NCHW.r.rxrN) rrKsplitrrrrrqjit is_scripting enumeraterr)r7rArrrrr intermediatesinfo stage_ends take_indices max_indexrlast_idxrfeat_idxrx_inters r:forward_intermediatesRDNet.forward_intermediates`s[&Y&D(DD& EIEVEVWEVTc$x...s3B78EV W"6s:"P /;<|!1 | <)  IIaLt(()A- 99 ! ! # #:,,L,,^i!m)rrrr@s r:forward_featuresRDNet.forward_featuress2 IIaL   a  MM! r< pre_logitscRU(aURUSS9$URU5$)NT)r)r)r7rArs r: forward_headRDNet.forward_heads$0:tyyty,L ! Lr<cJURU5nURU5nU$r>)rrr@s r:rB RDNet.forwards'  ! !! $   a r<) rrhrrrrrrrrr)r*avg)r`hrrr)r*r*r*r*r*r*r*)rrrQrQrQrQrQNTTFFFTra?rcrrr`?FTgelu layernorm2dNrbrbNNF)Tr>)NFFrF)rFT)rGrHrIrJrKrrrrboolrrrr1rqrignorerrr2rLrrrrrrrrBrMrNrOs@r:rrs#$9[!?^S%*^S+0^S!^S^S^S #^S#^S "!^S"#^S$S(]+%^S&'^S(uo)^S*+^S,"-^S^S@ YY   YY** YY 2C2hsm28<$$',1 ||1 eCcN341  1  1  1 !%1  tELL!5tELL7I)I#JJ K1 j./$# 3S >* $ M$Mr<rc[U[R5(a*[RR UR 5 g[U[R 5(aU[RRUR S5 [RRURS5 g[U[R5(a[RRURS5 U(aRSU;aKUR RRU5 URRRU5 gggg)Nrrzhead.) isinstancer2r4initkaiming_normal_weight BatchNorm2d constant_rLineardatamul_)rnamers r:rrs&"))$$  . FBNN + + &--+ &++q) FBII & & &++q) GtO MM   # #O 4 KK   ! !/ 2$4 'r<cSU;aU$SU;aUSn0nUR5Hup4URSS5nXBU'M U$)zRemap NV checkpoints -> timm z stem.0.weightmodelz stem.stem.zstem.)itemsreplace) state_dictrout_dictkvs r:checkpoint_filter_fnr!sZ*$*( H  " IIlG , # Or<c F[[X4[[SSS9S.UD6nU$)N)rrrr*T) out_indicesflatten_sequential)pretrained_filter_fn feature_cfg)rrr!r)variant pretrainedrrs r: _create_rdnetr)s4  w 1\dK    E Lr<c 8USSSSS[[SSSS S S S .UE$) Nr)r*rr)r)r)g?bicubiczstem.0zhead.fczarXiv:2403.19588z:DenseNets Reloaded: Paradigm Shift Beyond ResNets and ViTsz!https://github.com/naver-ai/rdnetz apache-2.0)urlr input_size pool_sizecrop_pct interpolationmeanstd first_conv classifier paper_ids paper_name origin_urllicenser )r,rs r:_cfgr9s<=v)%.Bi'R9    r<znaver-ai/rdnet_tiny.nv_in1k) hf_hub_idznaver-ai/rdnet_small.nv_in1kznaver-ai/rdnet_base.nv_in1kznaver-ai/rdnet_large.nv_in1kz(naver-ai/rdnet_large.nv_in1k_ft_in1k_384)r*r;) r<)r:r-r/r.)zrdnet_tiny.nv_in1kzrdnet_small.nv_in1kzrdnet_base.nv_in1kzrdnet_large.nv_in1kzrdnet_large.nv_in1k_ft_in1k_384c SnSS/S/-S/S--S/-S/U-SS S /S /-S /S--S /-S .n[SS U0[U40UD6D6nU$)Nr)r`rrrrr*rrrrQrrrrrrkr() rdnet_tinyr)rr(rn_layer model_argsrs r:r?r?sGu uqy0C583=L(+i7)+zlQ.>>*M J  \: \jA[TZA[ \E Lr<c SnSS/S/-S/US- --S/S--S/U-S S S /S /-S /US- --S /S--S .n[SSU0[U40UD6D6nU$)N Hr`rrrr* NTTFFFFFFTFrrrQr>r() rdnet_smallr@rAs r:rIrIsGu u! '<<uqyH3=h(+i7)+zlgk.JJj\\]M]] J  ]J ]$zB\U[B\ ]E Lr<c SnSS/S/-S/US- --S/S--S /U-S S S /S /-S /US- --S /S--S.n[SSU0[U40UD6D6nU$)NrEx`rriPrr*rHrrrQr>r() rdnet_baser@rAs r:rNrNsG u u! '<<uqyH3=h(+i7)+zlgk.JJj\\]M]] J  \: \jA[TZA[ \E Lr<c SnSS/S/-S/US- --S/S--S /U-S S S /S /-S /US- --S /S--S.n[SSU0[U40UD6D6nU$)Nr<rrihrr*) NTTFFFFFFFTFrrrQr>r() rdnet_larger@rAs r:rSrS%sG  1(== I3=o(+i7)+zlgk.JJj\\]M]] J  ]J ]$zB\U[B\ ]E Lr<)Nrr )r)5__doc__ functoolsrtypingrrrrrr rqtorch.nnr2 timm.datar r timm.layersr rrrrrrr_builderr _featuresr _manipulater _registryrr__all__rLrrQrr\r^r3rrrr!r)r9 default_cfgsr?rIrNrSror<r:r`sU ?? A222*+$< )BII2ryy473700f&&DGBIIGT 3   %/102/102'+< 3((D &          r<