RЦiI ZSrSSKJr SSKJrJrJrJrJr SSK r SSK J r SSK J r Jr SSKJrJrJrJrJrJr SSKJr SS KJr SS KJr SS KJrJr S /r"S S\ R@5r!"SS\ R@5r""SS\ R@5r#"SS\ R@5r$"SS\ R@5r%"SS \ R@5r&S'Sjr'\"\'"SS9\'"SS9\'"SS9\'"SSS9\'"SSSSS 9S!.5r(S(S"jr)\S(S#j5r*\S(S$j5r+\S(S%j5r,\S(S&j5r-g))z InceptionNeXt paper: https://arxiv.org/abs/2303.16900 Original implementation & weights from: https://github.com/sail-sg/inceptionnext )partial)ListOptionalTupleUnionTypeNIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD) trunc_normal_DropPathcalculate_drop_path_rates to_2tuple get_paddingSelectAdaptivePool2d)build_model_with_cfg)feature_take_indices)checkpoint_seq)register_modelgenerate_default_cfgsMetaNeXtc Z^\rSrSrSrS S\S\S\S\S\4 U4SjjjrS rS r U=r $) InceptionDWConv2dz$Inception depthwise convolution in_chssquare_kernel_sizeband_kernel_size branch_ratiodilationcd>XgS.n[T U]5 [X-5n [X%S9n [X5S9n [R "XU4XU S.UD6Ul[R "XSU44SU 4SU4U S.UD6Ul[R "XUS44U S4US4U S.UD6UlUSU -- XU 4Ul g)Ndevicedtype)r )paddingr groupsrr) super__init__intrnnConv2d dwconv_hwdwconv_wdwconv_h split_indexes) selfrrrrr r#r$ddgcsquare_padding band_padding __class__s Y/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/timm/models/inception_next.pyr)InceptionDWConv2d.__init__s /  & '$%7K"#3G  &H"bHDFH Q()P %H bPLNP  %q)P!1%1 bPLNP %q2vorr:c[R"XRSS9up#pE[R"UUR U5UR U5UR U54SS9$)Nr)dim)torchsplitr0catr-r.r/)r1xx_idx_hwx_wx_hs r7forwardInceptionDWConv2d.forward4s_${{1.@.@aHCyy  NN4 MM#  MM#      r9)r/r-r.r0)r' g?rNN) __name__ __module__ __qualname____firstlineno____doc__r*floatr)rD__static_attributes__ __classcell__r6s@r7rrs` '($&"';;!$;" ;  ;  ;;4  r9rc^\rSrSrSrSS\R SSSSS4S\S\\S\\S \ \RS \\ \RS \ S \ 4U4S jjjr SrSrU=r$)ConvMlp?zMLP using 1x1 convs that keeps spatial dims copied from timm: https://github.com/huggingface/pytorch-image-models/blob/v0.6.11/timm/models/layers/mlp.py NT in_featureshidden_features out_features act_layer norm_layerbiasdropc >XS.n [T U]5 U=(d UnU=(d Un[U5n[R"X4SUSS.U D6UlU(a U"U40U D6O[R "5UlU"5Ul[R"U5Ul [R"X#4SUSS.U D6Ul g)Nr"rr) kernel_sizerY) r(r)rr+r,fc1IdentitynormactDropoutrZfc2) r1rTrUrVrWrXrYrZr#r$r2r6s r7r)ConvMlp.__init__Ds / #2{ )8[99[]qtTUw]Z\]9CJ5"5 ;JJt$ 99_^PTUVPW^[]^r9cURU5nURU5nURU5nURU5nUR U5nU$N)r]r_r`rZrbr1r?s r7rDConvMlp.forward\sH HHQK IIaL HHQK IIaL HHQKr9)r`rZr]rbr_)rGrHrIrJrKr+ReLUr*rrModuleboolrLr)rDrMrNrOs@r7rQrQ?s.2*.)+48__&c]_#3- _ BII _ !bii1 ____0r9rQc^\rSrSrSrSSS\R \"\RSS9SS S S 4 S \ S \ S \ S\ S\ \RS\ \RS\ S\4U4SjjjrSS \ S \\ 4SjjrSS\4SjjrSrU=r$)MlpClassifierHeadezMLP classification head avgr'ư>)epsrSTNrT num_classes pool_type mlp_ratiorWrXrZrYc >XS.n [T U]5 SUlXl[ XA-5=Uln U(dS5e[ USS9Ul[R"XRR5-U 4SU0U D6Ul U"5Ul U"U 40U D6Ul [R"X4SU0U D6Ul[R"U5Ulg)Nr"FCannot disable poolingTrsflattenrY)r(r)use_convrTr* num_featuresr global_poolr+Linear feat_multr]r`r_rbrarZ)r1rTrrrsrtrWrXrZrYr#r$r2rUr6s r7r)MlpClassifierHead.__init__is /  &.1)2I.JJO222y/)TR99[+;+;+E+E+GGj_cjgij;5"5 99_KKKJJt$ r9cUbU(dS5e[USS9UlUS:a'[R"URU5Ulg[R "5Ulg)NrvTrwr)rr{r+r|rzr^rb)r1rrrss r7resetMlpClassifierHead.resetsT   66 693iQUVD @Ka299T.. <UWU`U`Ubr9 pre_logitscURU5nURU5nURU5nURU5nUR U5nU(aU$UR U5$re)r{r]r`r_rZrbr1r?rs r7rDMlpClassifierHead.forwards[   Q  HHQK HHQK IIaL IIaLq/DHHQK/r9) r`rZr]rbr{rTr_rzryreF)rGrHrIrJrKr+GELUr LayerNormr*strrLrrirjr)rrrDrMrNrOs@r7rlrles $" )+*1",,D*I%%% %  % BII %RYY%%%%8cc#c0T00r9rlc^\rSrSrSrS\\R\S\RSSSS4 S\ S \ S \ \RS \ \RS \ \RS \ S\ \RS\ S\ 4U4SjjjrSrSrU=r$) MetaNeXtBlockzMetaNeXtBlock 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. rrprSNr;r token_mixerrX mlp_layerrtrW ls_init_value drop_pathc x>XS.n [T U]5 U"U4SU0U D6UlU"U40U D6UlU"U[ Xa-54SU0U D6UlU(a.[ R"U[R"U40U D6-5OSUl U S:a[U 5Ul g[ R"5Ul g)Nr"r rWrS)r(r)rr_r*mlpr+ Parameterr<onesgammar r^r)r1r;r rrXrrtrWrrr#r$r2r6s r7r)MetaNeXtBlock.__init__s / &sDXDDs)b) S#io"6R)RrRLYR\\-%**S2GB2G"GH_c 09B),BKKMr9c UnURU5nURU5nURU5nURb-UR URR SSSS55nUR U5U-nU$)Nr)rr_rrmulreshaper)r1r?shortcuts r7rDMetaNeXtBlock.forwardsv   Q  IIaL HHQK :: !djj((B156A NN1  (r9)rrrr_r)rGrHrIrJrKrr+ BatchNorm2drQrr*rrirLr)rDrMrNrOs@r7rrs+<*,..)0 )+#'!RRRbii R RYY R BII RRBIIR!RRR,r9rc^\rSrSrSSSSS\\R SSSS4 S\S\S \S \S \\\4S \ \ \ S \ S\ \RS\ \RS\ \ \RS\ 4U4SjjjrSrSrU=r$) MetaNeXtStage)rrN?rrout_chsstridedepthr drop_path_ratesrrrWrXrtc>XS.n[TU]5 SUlUS:d USUS:waA[R"U "U40UD6[R "UU4SUUSS.UD65UlO[R"5UlU=(d S/U-n/n[U5H,nUR[S UUSUUUUU U U S.UD65 M. [R"U6Ul g) Nr"Frrr)r\rr rS)r;r rrrrWrXrt) r(r)grad_checkpointingr+ Sequentialr, downsampler^rangeappendrblocks)r1rrrrr rrrrWrXrtr#r$r2 stage_blocksir6s r7r)MetaNeXtStage.__init__s   / "' A:! 3 mm6(R( !"!%a[    DO!kkmDO)9bTE\ uA    !!!)!,+'#%# ! ! mm\2 r9cURU5nUR(a;[RR 5(d[ UR U5nU$UR U5nU$re)rrr<jit is_scriptingrrrfs r7rDMetaNeXtStage.forwardsV OOA   " "599+A+A+C+Ct{{A.A AAr9)rrr)rGrHrIrJrr+rr*rrrrLrrir)rDrMrNrOs@r7rrs (.59#&+<)+48 030303 03  03 CHo 03&d5k203!03bii03BII03!bii1030303dr9rc|^\rSrSrSrSSSSSS\\R\RS S S S S S 4S \ S\ S\ S\ S\ \ S4S\ \ S4S\ \ \R\\ \R4S\ \RS\ \RS\ \ \ \ S44S\S\S\4U4SjjjrSr\R(R*S2Sj5r\R(R*S\R4Sj5rS3S\ S\\ 4S jjr\R(R*S4S!j5r\R(R*S"5rS5S#\R8S$\\ \ \\ 4S%\S&\S'\ S(\S\ \\R8\ \R8\\R8444S)jjrS6S$\ \ \\ 4S*\S+\4S,jjrS-r S2S.\4S/jjr!S0r"S1r#U=r$$)7raXMetaNeXt 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. r'rnro r'r' r'`i)rrrr'rSrpNin_chansrrr{ output_stridedepths.dims token_mixersrXrW mlp_ratios drop_ratedrop_path_raterc>[TU]5 XS.n[U5n[U[[ 45(dU/U-n[U [[ 45(dU /U-n X lXlX0lXl /Ul [R"[R"XS4SSS.UD6U"US40UD65Ul[XSS9nUSnSnSn[R"5Ul[#U5HnUS:XdUS:aSOSnUU:a US:aUU-nSnUU-nUS ;aSOSnUUnUR R%['UU4US:aUOSUU4UUUUU U UUUU US . UD65 UnU=R[)UUS U3S 9/- sl M UUl[-UR*U4URU S .UD6UlUR.R*UlUR3UR45 g)Nr"rr)r\rT) stagewiserr)rr) rr rrrrWrrXrtzstages.)num_chs reductionmodule)rsrZ)r(r)len isinstancelisttuplerrrr{r feature_infor+rr,stemrstagesrrrdictrzrlheadhead_hidden_sizeapply _init_weights)r1rrrr{rrrrrXrWrrrrr#r$r2 num_stagedp_ratesprev_chs curr_strider rrfirst_dilationrr6s r7r)MetaNeXt.__init__ s2$  /K ,u 66(>I5L*tUm44$ 1J& &"MM IIhQ GQq GB G tAw %" %  -^tT7 mmo y!A%*a!eQFm+ F" 6 !K"*f"4Q!N1gG KK  } "#QvA((3Qi ( +#(O%$Q-   H   $x;Y`ab`cWd"e!f f /"0%%d&7&7wPTP`P`gpwtvw $ 6 6 4%%&r9c[U[R[R45(aM[ UR SS9 UR b+[RRUR S5 ggg)Ng{Gz?)stdr) rr+r,r|r weightrYinit constant_)r1ms r7rMetaNeXt._init_weightsSsV a"))RYY/ 0 0 !(( ,vv!!!!&&!," 1r9c0[SU(aSS9$SS/S9$)Nz^stemz^stages\.(\d+))z^stages\.(\d+)\.downsample)r)z^stages\.(\d+)\.blocks\.(\d+)N)rr)r)r1coarses r7 group_matcherMetaNeXt.group_matcherYs/(.$  685  r9returnc.URR$re)rrbr1s r7get_classifierMetaNeXt.get_classifiercsyy}}r9cFXlURRX5 gre)rrrr)r1rrr{s r7reset_classifierMetaNeXt.reset_classifiergs&  1r9c6URH nXlM gre)rr)r1enabless r7set_grad_checkpointingMetaNeXt.set_grad_checkpointingksA#) r9c[5$re)setrs r7no_weight_decayMetaNeXt.no_weight_decayps u r9r?indicesr_ stop_early output_fmtintermediates_onlycUS;dS5e/n[[UR5U5upURU5n[R R 5(dU(d URn OURSU S-n [U 5H%upU "U5nX;dMURU5 M' U(aU$X4$)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 Returns: )NCHWzOutput shape must be NCHW.Nr) rrrrr<rr enumerater) r1r?rr_rrr intermediates take_indices max_indexrfeat_idxstages r7forward_intermediatesMetaNeXt.forward_intermediatests*Y&D(DD& "6s4;;7G"Q  IIaL 99 ! ! # #:[[F[[)a-0F(0OHaA'$$Q' 1  r9 prune_norm prune_headc[[UR5U5upEURSUS-UlU(aURSS5 U$)z?Prune layers not required for specified intermediates. Nrrro)rrrr)r1rr r rrs r7prune_intermediate_layers"MetaNeXt.prune_intermediate_layerssK#7s4;;7G"Q kk.9q=1   ! !!U +r9cJURU5nURU5nU$re)rrrfs r7forward_featuresMetaNeXt.forward_featuress! IIaL KKNr9rcPU(aURXS9$URU5$)N)r)rrs r7 forward_headMetaNeXt.forward_heads#6@tyyy2RdiiPQlRr9cJURU5nURU5nU$re)rrrfs r7rDMetaNeXt.forwards'  ! !! $   a r9) rrr{rrrrrrzrrrre)T)NFFrF)rFT)%rGrHrIrJrKrr+rrr*rrrrrirrLr)rr<rignorerrrrrrTensorrjrr rrrDrMrNrOs@r7rrs&#$!#&2$7J[*,..)+6B!$&#'!E'E'E' E'  E' #s(O E'S/E' RYYd299o1F FGE'RYYE'BIIE'c5c?23E'E'"E'!E'E'N-  YY   YY 2C2hsm2 YY** YY 8<$$',( ||( eCcN34(  (  (  ( !%(  tELL!5tELL7I)I#JJ K( X./$# 3S >*    S$Sr9c 2USSSSS[[SSSS . UE$) Nrn)r'r)rg?bicubiczstem.0zhead.fc2z apache-2.0) urlrr input_size pool_sizecrop_pct interpolationmeanr first_conv classifierlicenser )rkwargss r7_cfgr&s3=vI%.Bj    r9ztimm/) hf_hub_idgffffff?)r'r)r'rr) r(r)r'rrr)zinception_next_atto.sail_in1kzinception_next_tiny.sail_in1kzinception_next_small.sail_in1kzinception_next_base.sail_in1kz!inception_next_base.sail_in1k_384c <[[X4S[SSS90UD6nU$)N feature_cfg)rrrr'T) out_indicesflatten_sequential)rrr)variant pretrainedr%models r7_create_inception_nextr0s1 ' \dK   E Lr9c `[SS[[SSS9S9n[SSU0[U40UD6D6$) N)rrr)(Pi@rg?)rrrrrr.)inception_next_atto)rrrr0r.r% model_argss r7r7r7sB"4.QUVJ " mJ mRVWaRlekRl mmr9c N[SS[S9n[SSU0[U40UD6D6$)Nrrr6r.)inception_next_tinyrrr0r8s r7r;r;s7"5&J " mJ mRVWaRlekRl mmr9c N[SS[S9n[SSU0[U40UD6D6$)Nr'r'r'rr6r.)inception_next_smallr<r8s r7r@r@s7#6&J " nZ nSWXbSmflSm nnr9c N[SS[S9n[SSU0[U40UD6D6$)Nr>)iir6r.)inception_next_baser<r8s r7rDrDs7#8&J " mJ mRVWaRlekRl mmr9)r).rK functoolsrtypingrrrrrr<torch.nnr+ timm.datar r timm.layersr r rrrr_builderr _featuresr _manipulater _registryrr__all__rirrQrlrrrr& default_cfgsr0r7r;r@rDrr9r7rQs[ 55 Axx*+'< ,&  & R#bii#L-0 -0`&BII&R9BII9x}ryy}@ %%)&&*&'+'&*& *. Hs*%& 4nnnnoonnr9