khC ddZddlmZddlmZmZmZmZddlZddl m Z ddl m Z m Z ddlmZmZmZmZmZddlmZdd lmZdd lmZdd lmZmZd gZGd de j<ZGdde j<Z Gdde j<Z!Gdde j<Z"Gdde j<Z#Gdd e j<Z$d'dZ%ee%de%de%de%dde%dddd d!Z&d(d"Z'ed(d#Z(ed(d$Z)ed(d%Z*ed(d&Z+y))z InceptionNeXt paper: https://arxiv.org/abs/2303.16900 Original implementation & weights from: https://github.com/sail-sg/inceptionnext )partial)ListOptionalTupleUnionNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD) trunc_normal_DropPath to_2tuple get_paddingSelectAdaptivePool2d)build_model_with_cfg)feature_take_indices)checkpoint_seq)register_modelgenerate_default_cfgsMetaNeXtc2eZdZdZ dfd ZdZxZS)InceptionDWConv2dz% Inception depthwise convolution clt |t||z}t||}t||}t j |||||||_t j ||d|fd|fd|f||_t j |||df|df|df||_|d|zz |||f|_ y)Ndilation)paddingrgroupsrr) super__init__intrnnConv2d dwconv_hwdwconv_wdwconv_h split_indexes) selfin_chssquare_kernel_sizeband_kernel_size branch_ratiorgcsquare_padding band_padding __class__s V/var/www/teggl/fontify/venv/lib/python3.12/site-packages/timm/models/inception_next.pyr zInceptionDWConv2d.__init__s  ,& '$%7(K"#3hG  &"XbB Q() %H bJ  %q)!1%1 bJ %q2vor2r:ctj||jd\}}}}tj||j ||j ||j |fdS)Nr)dim)torchsplitr'catr$r%r&)r(xx_idx_hwx_wx_hs r1forwardzInceptionDWConv2d.forward2se${{1d.@.@aHdCyy  NN4 MM#  MM#      r2)r g?r)__name__ __module__ __qualname____doc__r r= __classcell__r0s@r1rrs  ! ;0 r2rcJeZdZdZddej dddffd ZdZxZS)ConvMlpz MLP using 1x1 convs that keeps spatial dims copied from timm: https://github.com/huggingface/pytorch-image-models/blob/v0.6.11/timm/models/layers/mlp.py NTcft||xs|}|xs|}t|}tj||d|d|_|r||ntj |_||_tj||_ tj||d|d|_ y)Nrr) kernel_sizebias) rr r r"r#fc1IdentitynormactDropoutdropfc2) r( in_featureshidden_features out_features act_layer norm_layerrJrPr0s r1r zConvMlp.__init__Bs #2{ )8[99[/qtTUwW3=J/2;;= ;JJt$ 99_lPTUVPWXr2c|j|}|j|}|j|}|j|}|j |}|SN)rKrMrNrPrQr(r8s r1r=zConvMlp.forwardWsH HHQK IIaL HHQK IIaL HHQKr2) r?r@rArBr"ReLUr r=rCrDs@r1rFrF=s+!ggY*r2rFceZdZdZdddej eejdddffd Zdd e d e e fd Z dd e fdZxZS)MlpClassifierHeadz MLP classification head avgrư>)epsrGTc t |d|_||_t ||zx|_} |sJdt |d|_tj||jjz| ||_ ||_ || |_ tj| |||_tj||_y)NFCannot disable poolingT pool_typeflatten)rJ)rr use_convrRr! num_featuresr global_poolr"Linear feat_multrKrNrMrQrOrP) r(rR num_classesrd mlp_ratiorUrVrPrJrSr0s r1r zMlpClassifierHead.__init__ds  &.1)k2I.JJO222y/)TR99[4+;+;+E+E+GG_cd;/ 99_kEJJt$ r2rkrdc||sJdt|d|_|dkDr&tj|j||_ytj |_y)NrbTrcr)rrhr"rirgrLrQ)r(rkrds r1resetzMlpClassifierHead.reset}sT   66 693iQUVD @Ka299T.. <UWU`U`Ubr2 pre_logitsc|j|}|j|}|j|}|j|}|j |}|r|S|j |SrX)rhrKrNrMrPrQr(r8ros r1r=zMlpClassifierHead.forwards[   Q  HHQK HHQK IIaL IIaLq/DHHQK/r2rXF)r?r@rArBr"GELUr LayerNormr r!rstrrnboolr=rCrDs@r1r\r\`sY ggr||6%2cc#c0T0r2r\cbeZdZdZdeej edejddffd Z dZ xZ S) MetaNeXtBlockz MetaNeXtBlock Block Args: dim (int): Number of input channels. drop_path (float): Stochastic depth rate. Default: 0.0 ls_init_value (float): Init value for Layer Scale. Default: 1e-6. rr_rGc Xt |||||_|||_||t ||z||_|r+t j|tj|znd|_ | dkDrt| |_ yt j|_ y)Nr)rUrG)rr token_mixerrMr!mlpr" Parameterr5onesgammar rL drop_path) r(r4rr{rV mlp_layerrlrU ls_init_valuerr0s r1r zMetaNeXtBlock.__init__s &sX>sO S#i#o"6)LFSR\\-%**S/"ABY] 09B),BKKMr2c |}|j|}|j|}|j|}|j-|j |jj dddd}|j ||z}|S)Nr)r{rMr|rmulreshaper)r(r8shortcuts r1r=zMetaNeXtBlock.forwardsv   Q  IIaL HHQK :: !djj((B156A NN1  (r2) r?r@rArBrr" BatchNorm2drFrsr r=rCrDs@r1rxrxs5)~~ggR(r2rxc LeZdZdddddeej ddf fd ZdZxZS) MetaNeXtStage)rrN?ryc t|d|_|dkDs |d|dk7r=tj| |tj ||d||d|_ntj|_|xsdg|z}g} t|D]*} | jt||d|| ||| | | ,tj| |_ y)NFrrr)rIstriderrG)r4rrrr{rUrVrl) rr grad_checkpointingr" Sequentialr# downsamplerLrangeappendrxblocks)r(r)out_chsrdepthrdrop_path_ratesrr{rUrVrl stage_blocksir0s r1r zMetaNeXtStage.__init__s "' A:! 3 mm6"  !!%a[  DO!kkmDO)9bTE\ u A    !!)!,+'#%# !  mm\2 r2c|j|}|jr6tjj st |j |}|S|j |}|SrX)rrr5jit is_scriptingrrrYs r1r=zMetaNeXtStage.forwardsS OOA   " "599+A+A+Ct{{A.A AAr2) r?r@rArr"rsr r=rCrDs@r1rrs/  )gg+3Zr2rc~eZdZdZddddddeej ejdd d d f fd Zd Z e jjd#d Z e jjdejfdZd$dedeefdZe jjd%dZe jjdZ d&de j,deeeeefdededededeee j,ee j,ee j,fffdZ d'deeeefdedefdZdZd#d efd!Zd"ZxZ S)(ra MetaNeXt A PyTorch impl of : `InceptionNeXt: When Inception Meets ConvNeXt` - https://arxiv.org/abs/2303.16900 Args: in_chans (int): Number of input image channels. Default: 3 num_classes (int): Number of classes for classification head. Default: 1000 depths (tuple(int)): Number of blocks at each stage. Default: (3, 3, 9, 3) dims (tuple(int)): Feature dimension at each stage. Default: (96, 192, 384, 768) token_mixers: Token mixer function. Default: nn.Identity norm_layer: Normalization layer. Default: nn.BatchNorm2d act_layer: Activation function for MLP. Default: nn.GELU mlp_ratios (int or tuple(int)): MLP ratios. Default: (4, 4, 4, 3) drop_rate (float): Head dropout rate drop_path_rate (float): Stochastic depth rate. Default: 0. ls_init_value (float): Init value for Layer Scale. Default: 1e-6. rr]r^ rr r`i)ryryryrrGr_c`t|t|}t|tt fs|g|z}t| tt fs| g|z} ||_||_| |_g|_ tjtj||ddd||d|_ tjd| t!|j#|Dcgc]}|j%}}|d}d}d}tj|_t)|D]}|dk(s|dkDrdnd}||k\r |dkDr||z}d}||z}|dvrdnd}||}|j&j+t-|||dkDr|nd||f||||| | |||| | |}|xjt/||d| gz c_ ||_t3|j0||j| |_|j4j0|_|j9|j:ycc}w) Nrry)rIrrr)rr) rrrrrrUr{rVrlzstages.)num_chs reductionmodule)rdrP)rr len isinstancelisttuplerkrh drop_rate feature_infor"rr#stemr5linspacesumr6toliststagesrrrdictrgr\headhead_hidden_sizeapply _init_weights)r(in_chansrkrh output_stridedepthsdims token_mixersrVrU mlp_ratiosrdrop_path_rater num_stager8dp_ratesprev_chs curr_striderrrfirst_dilationrr0s r1r zMetaNeXt.__init__s8 K ,u 6(>I5L*tUm4$ 1J&&"MM IIhQQq A tAw   ).q.#f+(V(\(\]c(de1AHHJee7 mmo y! gA%*a!eQFm+ F" 6 !K"*f"4Q!N1gG KK  }!"QvA((3Qi ( +#(O%$Q-  H   $x;Y`ab`cWd"e!f f - g.%%d&7&7PTP`P`gpq $ 6 6 4%%&AfsH+ct|tjtjfrOt |j d|j +tjj|j dyyy)Ng{Gz?)stdr) rr"r#rir weightrJinit constant_)r(ms r1rzMetaNeXt._init_weights@sS a"))RYY/ 0 !(( ,vv!!!!&&!," 1r2c2td|rdSddgS)Nz^stemz^stages\.(\d+))z^stages\.(\d+)\.downsample)r)z^stages\.(\d+)\.blocks\.(\d+)N)rr)r)r(coarses r1 group_matcherzMetaNeXt.group_matcherFs/(.$  685  r2returnc.|jjSrX)rrQr(s r1get_classifierzMetaNeXt.get_classifierPsyy}}r2rkrhcJ||_|jj||yrX)rkrrn)r(rkrhs r1reset_classifierzMetaNeXt.reset_classifierTs&  [1r2c4|jD] }||_ yrX)rr)r(enabless r1set_grad_checkpointingzMetaNeXt.set_grad_checkpointingXs *A#)A  *r2ctSrX)setrs r1no_weight_decayzMetaNeXt.no_weight_decay]s u r2r8indicesrM stop_early output_fmtintermediates_onlycr|dvsJdg}tt|j|\}} |j|}tj j s|s |j} n|jd| dz} t| D]#\} } | |}| |vs|j|%|r|S||fS)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 Returns: )NCHWzOutput shape must be NCHW.Nr) rrrrr5rr enumerater) r(r8rrMrrr intermediates take_indices max_indexrfeat_idxstages r1forward_intermediateszMetaNeXt.forward_intermediatesas*Y&D(DD& "6s4;;7G"Q i IIaL 99 ! ! #:[[F[[)a-0F(0 (OHeaA<'$$Q' (  -r2 prune_norm prune_headctt|j|\}}|jd|dz|_|r|jdd|S)z@ Prune layers not required for specified intermediates. Nrrr^)rrrr)r(rrrrrs r1prune_intermediate_layersz"MetaNeXt.prune_intermediate_layerssM#7s4;;7G"Q ikk.9q=1   ! !!U +r2cJ|j|}|j|}|SrX)rrrYs r1forward_featureszMetaNeXt.forward_featuress! IIaL KKNr2rocN|r|j||S|j|S)N)ro)rrqs r1 forward_headzMetaNeXt.forward_heads%6@tyyzy2RdiiPQlRr2cJ|j|}|j|}|SrX)rrrYs r1r=zMetaNeXt.forwards'  ! !! $   a r2rrrX)T)NFFrF)rFT)!r?r@rArBrr"rrsr rr5rignorerModulerr!rrurrrTensorrrrvrrrrrr=rCrDs@r1rrs&$*~~gg#A'F-  YY   YY 2C2hsm2 YY** YY 8<$$',( ||( eCcN34(  (  (  ( !%(  tELL!5tELL7I)I#JJ K( X./$# 3S >*    S$Sr2c 0|dddddttddd |S) Nr])rr)rg?bicubiczstem.0zhead.fc2) urlrk input_size pool_sizecrop_pct interpolationmeanr first_conv classifierr)rkwargss r1_cfgrs0=vI%.Bj     r2ztimm/) hf_hub_idgffffff?)rr)rrr) rr)rrrr)zinception_next_atto.sail_in1kzinception_next_tiny.sail_in1kzinception_next_small.sail_in1kzinception_next_base.sail_in1kz!inception_next_base.sail_in1k_384c Dtt||fdtddi|}|S)N feature_cfg)rrrrT) out_indicesflatten_sequential)rrr)variant pretrainedrmodels r1_create_inception_nextr s3 ': \dK   E Lr2c ltddttdd}tdd|it|fi|S) N)rrr)(Pi@rg?)r+r,rrrr )inception_next_atto)rrrr r r model_argss r1rrsB"4.QUVJ " mJ mRVWaRlekRl mmr2c Ttddt}tdd|it|fi|S)Nrrrr )inception_next_tinyrrr rs r1rrs7"5&J " mJ mRVWaRlekRl mmr2c Ttddt}tdd|it|fi|S)Nrrrrrr )inception_next_smallrrs r1rrs7#6&J " nZ nSWXbSmflSm nnr2c Ttddt}tdd|it|fi|S)Nr)iirr )inception_next_baserrs r1rrs7#8&J " mJ mRVWaRlekRl mmr2)rr),rB functoolsrtypingrrrrr5torch.nnr" timm.datar r timm.layersr r r rr_builderr _featuresr _manipulater _registryrr__all__rrrFr\rxrrr default_cfgsr rrrrr2r1r-s[ // A]]*+'< ,$  $ N bii F*0 *0Z$BII$N4BII4nyryyyx%%)&&*&'+'&*& *. Hs*%& 4nnnnoonnr2