khMM ^dZddlmZddlmZmZmZmZmZddl Z ddl m Z ddl m Z mZddlmZmZmZmZmZmZmZddlmZdd lmZdd lmZdd lmZmZd gZ Gd de jBZ"Gdde jBZ#de$fdZ%Gdde jBZ&Gdde jNZ(Gdd e jBZ)d+dZ*dZ+d,dZ,d-dZ-ee-de-de-d e-d!e-d"d#dd$%d&Z.ed,d'Z/ed,d(Z0ed,d)Z1ed,d*Z2y).z? RDNet Copyright (c) 2024-present NAVER Cloud Corp. Apache-2.0 )partial)ListOptionalTupleUnionCallableNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)DropPathNormMlpClassifierHeadClassifierHeadEffectiveSEModulemake_divisible get_act_layerget_norm_layer)build_model_with_cfg)feature_take_indices) named_apply)register_modelgenerate_default_cfgsRDNetc$eZdZfdZdZxZS)Blockc t|tjtj|||ddd||tj||ddd|tj||ddd|_yNr)groups kernel_sizestridepaddingrr!r"r#)super__init__nn SequentialConv2dlayersselfin_chs inter_chsout_chs norm_layer act_layer __class__s M/var/www/teggl/fontify/venv/lib/python3.12/site-packages/timm/models/rdnet.pyr&zBlock.__init__sj mm IIffV1VW X v  IIfiQq! L K IIia1 M   c$|j|SNr*r,xs r3forwardz Block.forward#{{1~r4__name__ __module__ __qualname__r&r: __classcell__r2s@r3rrs  r4rc$eZdZfdZdZxZS)BlockESEc t|tjtj|||ddd||tj||ddd|tj||dddt ||_yr)r%r&r'r(r)rr*r+s r3r&zBlockESE.__init__(sr mm IIffV1VW X v  IIfiQq! L K IIia1 M g &   r4c$|j|Sr6r7r8s r3r:zBlockESE.forward3r;r4r<rAs@r3rCrC's   r4rCblockc|jj}|dk(rtS|dk(rtSJd|d)NrFblockesezUnknown block type (z).)lowerstriprrC)rFs r3_get_block_typerK7sD KKM   !E  * 6,UG266ur4ceZdZ ddedededededededed ed ed effd Zd eejdejfdZ xZ S) DenseBlocknum_input_features growth_ratebottleneck_width_ratiodrop_path_rate drop_raterand_gather_step_prob block_idx block_type ls_init_valuer0r1c jt |||_||_||_||_||_| dkDr+tj| tj|znd|_ t|}t||zdz dz} t||_t||| || | |_y)Nr)r-r.r/r0r1)r%r&rRrQrSrTrOr' Parametertorchonesgammaintr drop_pathrKr*)r,rNrOrPrQrRrSrTrUrVr0r1r.r2s r3r&zDenseBlock.__init__Bs ",%:""&N[^_N_R\\-%**[2I"IJei +& *-CCaGH1L !.11oj1%!   r4r9returnctj|d}|j|}|j-|j |jj dddd}|j |}|S)Nr)rZcatr*r\mulreshaper^r8s r3r:zDenseBlock.forwardes^ IIaO KKN :: !djj((B156A NN1 r4) @re@rgrgrrư> layernorm2dgelu) r=r>r?r]floatstrr&rrZTensorr:r@rAs@r3rMrMAs')!,/$'"+.%#'+#! #! ! %* ! " !  ! $)! ! ! !! ! ! Fell+ r4rMcVeZdZfdZdej dej fdZxZS) DenseStagec t|t|D]/}td|||||d|}||z }|j d||1||_y)N)rNrOrQrT dense_block)r%r&rangerM add_modulenum_out_features) r, num_blockrNdrop_path_ratesrOkwargsilayerr2s r3r&zDenseStage.__init__qsx y! 6A#5'.q1   E + -  OOk!-u 5 6!3r4 init_featurer_ct|g}|D]}||}|j|tj|dS)Nr)appendrZrb)r,r{featuresmodule new_features r3r:zDenseStage.forwardsA > )F *K OOK ( )yy1%%r4)r=r>r?r&rZrmr:r@rAs@r3rorops# 3&ELL&U\\&r4roc+eZdZ d*dedededeeeeefdeeeeefdeeeeefdeeeeefde d e d e d ed ed ede dededeee fdede e de de f*fd Z e jjd+dZe jjd,dZe jjdej&fdZd-dede efdZ d.de j,de eeeefdeded ed!edeee j,ee j,ee j,fffd"Z d/deeeefd#ed$efd%Zd&Zd+d'efd(Zd)ZxZS)0rin_chans num_classes global_pool growth_ratesnum_blocks_listrUis_downsample_blockrPtransition_compression_ratiorV stem_type patch_sizenum_init_featureshead_init_scalehead_norm_first conv_biasr1r0norm_epsrRrQc "t#|t|t|cxk(rt|k(sJJt|}t |}| t ||}||_||_| dvsJ| dk(r|_tA|j4|||j||_tCt tD||ycc}w)a  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. N)eps)patchoverlapoverlap_tieredr)r!r"biastieredrr)r!r"r#rrrXr$) rvrNrOrPrRrwrVrUr0r1z dense_stages.)num_chs reductionrrO) pool_typerR)rrRr0)r)#r%r&lenrrrrrRr'r(r)stemr feature_info num_stagesrZlinspacesumsplittolistrsr]r}rodict dense_stages num_featureshead_hidden_sizenorm_prerheadIdentityr r _init_weights)$r,rrrrrrUrrPrrVrrrrrrr1r0rrRrQ stem_stridemid_chs curr_striderr9dp_ratesrrydense_stage_layerscompressed_num_featuresk_sizer"stager2s$ r3r&zRDNet.__init__s^ < C$8TC@S\}}||}|| vs|r|| k(r|j|}n|}|j|@|r|S| k(r|j|}||fScc}wcc} w)a Forward features that returns intermediates. Args: x: Input image tensor indices: Take last n blocks if int, all if None, select matching indices if sequence norm: Apply norm layer to compatible intermediates stop_early: Stop iterating over blocks when last desired intermediate hit output_fmt: Shape of intermediate feature outputs intermediates_only: Only return intermediate features )NCHWzOutput shape must be NCHW.r.rarN) rr]rrrrrrZjit is_scripting enumeraterr})r,r9rrrrr intermediatesinfo stage_ends take_indices max_indexrylast_idxrfeat_idxrx_inters r3forward_intermediateszRDNet.forward_intermediates1se&Y&D(DD& EIEVEVWTc$x...s3B78W W"6s:"P i/;r?r]rlrrrboolrkrrr&rZrignorerrr'Modulerrrmrrrrr:r@rAs@r3rrs#$9[!?RS%*RS+0RS!RSRSRS #RS#RS "!RS"#RS$S(]+%RS&'RS(uo)RS*+RS,"-RSh YY   YY** YY 2C2hsm28<$$',1 ||1 eCcN341  1  1  1 !%1  tELL!5tELL7I)I#JJ K1 j./$# 3S >* $ M$Mr4rct|tjr*tjj |j yt|tj rUtjj|j dtjj|jdyt|tjr}tjj|jd|rPd|vrK|j jj||jjj|yyyy)Nrrzhead.) isinstancer'r)initkaiming_normal_weight BatchNorm2d constant_rLineardatamul_)rnamers r3rrs&"))$  . FBNN + &--+ &++q) FBII & &++q) GtO MM   # #O 4 KK   ! !/ 2$4 'r4cd|vr|Sd|vr|d}i}|jD]\}}|jdd}|||<|S)z Remap NV checkpoints -> timm z stem.0.weightmodelz stem.stem.zstem.)itemsreplace) state_dictrout_dictkvs r3checkpoint_filter_fnrsc*$*( H  "1 IIlG ,  Or4c Ntt||fttddd|}|S)N)rrrrT) out_indicesflatten_sequential)pretrained_filter_fn feature_cfg)rrrr)variant pretrainedrxrs r3 _create_rdnetr s6  w  1\dK    E Lr4c 6|dddddttdddd d d |S) Nr)rrr)rrg?bicubiczstem.0zhead.fczarXiv:2403.19588z:DenseNets Reloaded: Paradigm Shift Beyond ResNets and ViTsz!https://github.com/naver-ai/rdnet) urlr input_size pool_sizecrop_pct interpolationmeanstd first_conv classifier paper_ids paper_name origin_urlr )r rxs r3_cfgrs9=v)%.Bi'R9    r4znaver-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)rr) r)rr rr)zrdnet_tiny.nv_in1kzrdnet_small.nv_in1kzrdnet_base.nv_in1kzrdnet_large.nv_in1kzrdnet_large.nv_in1k_ft_in1k_384c d}ddgdgzdgdzzdgzdg|zdd d gd gzd gdzzd gzd }tdd |it|fi|}|S)NrrerrrrrrrrrCrrrrrrUr) rdnet_tinyr rrrxn_layer model_argsrs r3rrsGu uqy0C583=L(+i7)+zlQ.>>*M J  \: \jA[TZA[ \E Lr4c d}ddgdgzdg|dz zzdgdzzdg|zd d d gd gzd g|dz zzd gdzzd }tdd|it|fi|}|S)N Hrerrrr NTTFFFFFFTFrrrCrr) rdnet_smallrr s r3r(r(sGu u! '<<uqyH3=h(+i7)+zlgk.JJj\\]M]] J  ]J ]$zB\U[B\ ]E Lr4c d}ddgdgzdg|dz zzdgdzzd g|zd d d gd gzd g|dz zzd gdzzd}tdd|it|fi|}|S)Nr$x`rriPrrr'rrrCrr) rdnet_baserr s r3r-r-sG u u! '<<uqyH3=h(+i7)+zlgk.JJj\\]M]] J  \: \jA[TZA[ \E Lr4c d}ddgdgzdg|dz zzdgdzzd g|zd d d gd gzd g|dz zzd gdzzd}tdd|it|fi|}|S)Nrrrihrr) NTTFFFFFFFTFrrrCrr) rdnet_largerr s r3r2r2sG  1(== I3=o(+i7)+zlgk.JJj\\]M]] J  ]J ]$zB\U[B\ ]E Lr4)Nrr)r)3__doc__ functoolsrtypingrrrrrrZtorch.nnr' timm.datar r timm.layersr r rrrrr_builderr _featuresr _manipulater _registryrr__all__rrrCrlrKrMr(rorrrr r default_cfgsrr(r-r2rrr4r3r?sT 99 A222*+$< ) BII  ryy  737,,^&&.{BII{| 3   %/102/102'+< 3((D &          r4