kh| ddlmZddlmZmZmZmZmZmZddl Z ddl m Z ddl m cm ZddlmZmZddlmZmZmZmZmZmZmZmZmZddlmZddlmZm Z dd l!m"Z"m#Z#m$Z$m%Z%m&Z&m'Z'dd l(m)Z)dd l*m+Z+dd l,m-Z-m.Z.dd l/m0Z0m1Z1ddgZ2e+Gdde jfZ4Gdde jfZ5Gdde jfZ6d&de7de8de6fdZ9d&de7de8de6fdZ: d'de7de;de8de8de6f dZe1d&de8de6fd#Z?e1d&de8de5fd$Z@e1d&de8de5fd%ZAy)))partial)CallableListOptionalSequenceTupleUnionNIMAGENET_INCEPTION_MEANIMAGENET_INCEPTION_STD) SelectAdaptivePool2dLinear LayerTypePadType RmsNorm2d ConvNormAct create_conv2dget_norm_act_layer to_2tuple)build_model_with_cfg) SqueezeExciteUniversalInvertedResidual) BlockArgsEfficientNetBuilderdecode_arch_defefficientnet_init_weightsround_channelsresolve_act_layer)feature_take_indices)register_notrace_module)checkpoint_seq checkpoint)generate_default_cfgsregister_model MobileNetV5MobileNetV5EncoderceZdZdZ ddeeeefdedededede ede d e e d e e ffd Z d ee jd e jfdZxZS)"MobileNetV5MultiScaleFusionAdapteraMulti-layer fusion token adapter. Args: in_chs: List of input channel counts for each feature scale. out_chs: The number of output channels. output_resolution: The output resolution. expansion_ratio: The FFN expansion ratio. interpolation_mode: The upsampling interpolation mode. layer_scale_init_value: The initial value of the layer scale, no layer scale if None. in_chsout_chsoutput_resolutionexpansion_ratiointerpolation_modelayer_scale_init_valuenoskip act_layer norm_layerc t |t|tr t |n||_||_t||_||_ ||_ ||_ ||_ |xstj}| xst} t!|j |j d|j|| |j|j|_| |j |_y)Nr)r*r+dw_kernel_size_mid exp_ratior1r2r0r/)super__init__ isinstancersum in_channels out_channelsrr,r-r.r/r0nnGELUrrffnnorm) selfr*r+r,r-r.r/r0r1r2 __class__s S/var/www/teggl/fontify/venv/lib/python3.12/site-packages/timm/models/mobilenetv5.pyr7z+MobileNetV5MultiScaleFusionAdapter.__init__&s G&0&Bs6{DD&'89D*D0D"8DDK$RWWI(yJ(!!&&{{#:: DH4,,-DIinputsreturnc|djdd}g}t|D]]\}}|jdd}|d|dks |d|dkr"tj|||j}|j |_t j|d}|j|}|d|jdk7s|d|jdk7r|d|jdzdk7s|d|jdzdk7r#tj||jd}nF|d|jdz}|d|jdz} tj||| f|| f}|j|}|S)Nrr)sizemode)dimbilinear) kernel_sizestride) shape enumerateF interpolater.appendtorchcatr>r, avg_pool2dr?) r@rDhigh_resolutionresized_inputs_img feat_sizechannel_cat_imgs h_strides w_stridess rBforwardz*MobileNetV5MultiScaleFusionAdapter.forwardJsQioobc*ONF##3IIbcN Qr&z MobiletNet-V5 FT)rGNavg block_args num_classesin_chans stem_sizefix_stem num_featurespad_typeuse_msfa msfa_indicesmsfa_output_resolutionr1r2aa_layerse_layer se_from_exp round_chs_fn drop_ratedrop_path_rater/ global_poolc t|| xstj} | xst} t | | }|xst }||_||_d|_ | |_ | |_ |s||}t||dd|| | |_ td|||| | | ||| }tj||||_|j"|_|j$Dcgc]}|d c}|_|j(|_|r|x|_|_t/t1|j$|jd|_ t3|jDcgc]}|j$|d c}|_t7|j4||j| | |_t;| |_d |_d |_ n|j(|_||_d |_t;| |_|j*|j<jCz}tE||j,d ||_||j,|_ |rtjFd ntjH|_%|dkDrtM|j,|ntjH|_'tQ|y cc}wcc}w)a Args: block_args: Arguments for blocks of the network. num_classes: Number of classes for classification head. in_chans: Number of input image channels. stem_size: Number of output channels of the initial stem convolution. fix_stem: If True, don't scale stem by round_chs_fn. num_features: Number of output channels of the conv head layer. head_bias: If True, add a learnable bias to the conv head layer. pad_type: Type of padding to use for convolution layers. act_layer: Type of activation layer. norm_layer: Type of normalization layer. aa_layer: Type of anti-aliasing layer. se_layer: Type of Squeeze-and-Excite layer. se_from_exp: If True, calculate SE channel reduction from expanded mid channels. round_chs_fn: Callable to round number of filters based on depth multiplier. drop_rate: Dropout rate. drop_path_rate: Stochastic depth rate. layer_scale_init_value: Enable layer scale on compatible blocks if not None. global_pool: Type of pooling to use for global pooling features of the FC head. FrmrLrMpaddingr2r1 output_strider{rrr1r2rrrr/stagernum_chsr*r+r,r2r1 pool_typeNr)r))r6r7r<r=rrrrvrgrad_checkpointingr}r~r conv_stemr Sequentialblocksfeatures feature_info stage_endsr*rzhead_hidden_sizer lenr9 msfa_in_chsr)msfar r conv_head norm_head feat_multrFlattenIdentityflattenr classifierr)r@rurvrwrxryrzr{r|r}r~r1r2rrrrrrr/rnorm_act_layerbuilderfminum_pooled_chsrAs rBr7zMobileNetV5.__init__psX ( ,9 +J B,}&""'(&<#$Y/I$  ! &%#!)#9  mmWY %CD #,,/3/@/@A!1W:A#NN 8D DD  5 4S9J9J5KTM^M^ _`a bD "tO`O`#aD$5$5b$9)$D#abD :''$"&"="=%# DI 4kJD !DN!DN 'D $0D !DI3kJD !..1A1A1K1K1MMN*>4;P;PRS]efDN+D,A,ABDN(3rzz!} HSVW&!6!6 D]_]h]h]j!$'CB$bs  J=Kc|j|jg}|j|j|j |j |j |j |j |j|j |j|jtjtj|j|jgtj|SN)rbn1extendrrRrrrr<rDropoutrrr)r@layerss rB as_sequentialzMobileNetV5.as_sequentials..$((+ dkk" d&&' >> % MM$.. ) >> % MM$.. ) rzz|RZZ%?QR}}f%%rCcoarsec.td|rdSdS)Nz^conv_stem|bn1z^blocks\.(\d+)z^blocks\.(\d+)\.(\d+))stemr)dict)r@rs rB group_matcherzMobileNetV5.group_matchers%"(.$  4L  rCenablec||_yr)r)r@rs rBset_grad_checkpointingz"MobileNetV5.set_grad_checkpointings "(rCrEc|jSr)r)r@s rBget_classifierzMobileNetV5.get_classifiers rCc||_t||_|rtjdntj |_|dkDrt|j||_ ytj |_ y)Nrrr) rvr rr<rrrrrr)r@rvrs rBreset_classifierzMobileNetV5.reset_classifiers[&/+F(3rzz!} HSVW&!6!6 D]_]h]h]jrCxindicesr? stop_early output_fmtintermediates_only extra_blockscR|dvsJd|r |sJdg}|r&tt|jdz|\} } nMtt|j|\} } | D cgc]} |j| } } |j| } d} |j |}| | vr|j |t jjs|s |j} n|jd| } | D]%}| dz } ||}| | vs|j |'|r|S||fScc} w)aa 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 extra_blocks: Include outputs of all blocks and head conv in output, does not align with feature_info Returns: NCHWOutput shape must be NCHW./Must use intermediates_only for early stopping.rrN) r rrrrrRrSjit is_scripting)r@rrr?rrrr intermediates take_indices max_indexifeat_idxrblks rBforward_intermediatesz!MobileNetV5.forward_intermediatessB.Y&D(DD& % X'X X% &:3t{{;Ka;OQX&Y #L)&:3t;OQX&Y #L)8DE1DOOA.ELE 2I NN1  | #   # 99 ! ! #:[[F[[),F (C MHAA<'$$Q'  (  -;Fs'D$ prune_norm prune_headcp|r&tt|jdz|\}}n1tt|j|\}}|j|}|jd||_|t|jkrd|_d|_|r d|_d|_|j dd|S)z@ Prune layers not required for specified intermediates. rNrrp)r rrrrrr)r@rrrrrrs rBprune_intermediate_layersz%MobileNetV5.prune_intermediate_layersEs &:3t{{;Ka;OQX&Y #L)&:3t;OQX&Y #L) 2Ikk*9- s4;;' '!DN!DN !DN!DN  ! !!R (rCc|jd}g}|j|}||jvr|j||jD]/}|dz }||}||jvs|j|1|j|}|S|j|}|j r8t jjst|j|d}|S|j |}|S)NrrT)r) rrr}rRrrrSrrr"r@rrrrs rBforward_featureszMobileNetV5.forward_features]s 99 HMq!A4,,,$$Q'{{ ,A Ft000!((+  ,  -(A q!A&&uyy/E/E/G"4;;4@KKNrC pre_logitscZ|j|}|j|j|}|j|j|}|j|}|jdkDr,t j ||j|j}|r|S|j|S)Nrs)ptraining) rrrrrrPdropoutrr)r@rrs rB forward_headzMobileNetV5.forward_headts   Q  >> %q!A >> %q!A LLO >>B  !t~~ FA Hq!!rCcJ|j|}|j|}|Sr)rrr@rs rBr^zMobileNetV5.forwards'  ! !! $   a rCF)T)rtNFFrFF)rFTF)"r`rarbrcrrrdrgrfrrrrrer7rrSrignorerrr<Modulerrrhr rrrrrr^rirjs@rBr&r&ls $" $!&2*,-1.2,0,0 $%3!$&6:$+t(!t(t( t(  t(  t(t(t(t(s)t(%(t( *t(!+t(y)t(y)t( !t("##t($%t(&"'t((%-UO)t(*+t(l & YY D   YY)T)) YY kCkck8<$$',!&< ||< eCcN34<  <  <  < !%< <  tELL!5tELL7I)I#JJ K< @./$#!& 3S >*   0%,,5<<. "ell " " "%,,rCc" eZdZdZddddddddddd ed d dfd ed ed ededede edede e de e de e de e dede de de de e f fd Z d(dej de eeeefdeded ed!ed"ed#eeej eej eej fffd$Zdej d#ej fd%Zdej d#ej fd&Zdej d#ej fd'ZxZS))r'zMobileNetV5 Vision Encoderrm@Frp)rGrrrnNTrsrurwrxryr{r}r~r1r2rrrrrrr/c "t||xstj}| xst} | xst } d|_||_d|_|s| |}t||dd|| ||_ td|| | || | | || }tj||||_ |j|_|jDcgc]}|d c}|_d x|_|_t't)|j|d|_t-|j*Dcgc]}|j|d c}|_||_t3|j.|j"|j0| | |_t7|ycc}wcc}w) NrFrmrrrrrrorr)r6r7r<r=rrrvrrrrrrrrrrrzrr rr}r9rr~r)rr)r@rurwrxryr{r}r~r1r2rrrrrrr/rrrrAs rBr7zMobileNetV5Encoder.__init__s& ( ,9 ,}""'$Y/I$  ! &%#!)#9  mmWY %CD #,,/3/@/@A!1W:A488D10T5F5F1GVWXY4K\K\]R 1 1" 5i @]^&<#6##%%"99!    "$'!B ^s FF rrr?rrrrrEc~|dvsJd|r |sJdg}g} |r&tt|jdz|\} } nMtt|j|\} } | D cgc]} |j| } } |j| } d} |j |}| | vr|j || |j vr| j |tjjs|s |j}n|jd| }|D]D}| dz } ||}| | vr|j || |j vs4| j |F|r|S|j| |fScc} w)al 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: (Unused) Applies 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 extra_blocks: Include outputs of all blocks and head conv in output, does not align with feature_info Returns: rrrrrN) r rrrrrRr}rSrrr)r@rrr?rrrrrmsfa_intermediatesrrrrrrs rBrz(MobileNetV5Encoder.forward_intermediatess. Y&D(DD& % X'X X%   &:3t{{;Ka;OQX&Y #L)&:3t;OQX&Y #L)8DE1DOOA.ELE 2I NN1  | #   # t(( (  % %a ( 99 ! ! #:[[F[[),F -C MHAA<'$$Q'4,,,"))!,  -  yy+,m;;7Fs*E4cd}g}|j|}||jvr|j||jD]/}|dz }||}||jvs|j|1|j |S)Nrr)rr}rRrrrs rBrz#MobileNetV5Encoder.forward_featuress NN1  t(( (   #;; (C MHAA4,,,$$Q'  (yy''rCctd)Nz=MobileNetV5Encoder does not support classification use cases.)NotImplementedErrorrs rBrzMobileNetV5Encoder.forward_head(s!"abbrCc$|j|Sr)rrs rBr^zMobileNetV5Encoder.forward+s$$Q''rCr)r`rarbrcrrrdrgrfrrrrrer7rSrhr rrrrrr^rirjs@rBr'r's $ "*2*,-1.2,0,0 $%3!$&6:#F(!F(F( F(  F(  F(#3-F(%(F( *F(!+F(y)F(y)F(F(#F(F( "!F("%-UO#F(V8<$$',!&C<||C<eCcN34C< C<  C<  C<!%C<C< tELL!5tELL7I)I#JJ KC)rurxryrr2r1r/) rrrrrr<r=rr) rrrrrrarch_def model_kwargsrs rB_gen_mobilenet_v5rTsW$    & G4% p   & P( qa% F"8 C#c)^8JK''#L //L$WjILI LWjALA LrCurlc 0|dddddttddd |S) Nrl)rmr")rnrnrbicubiczconv_stem.convr) r rv input_size pool_sizecrop_pct interpolationmeanstd first_convrr )r rs rB_cfgr+s04}S[)'0F&l     rC)rmr,)r$rvrl)rv)mobilenetv5_300m_encmobilenetv5_300mzmobilenetv5_base.untrainedc J|jdd}t d|d|d|}|S)zMobileNet V5 Vision Encoderr{sameT)rrr{)r-)rr)rrr{rs rBr-r-#s@zz*f-H       E LrCc tdd|i|}|S)Nr)r.rrrrs rBr.r.1  RZ R6 RE LrCc tdd|i|}|S)Nr)rr2r3s rBrr7r4rCr)rNFF)rp)B functoolsrtypingrrrrrr rStorch.nnr<torch.nn.functional functionalrP timm.datar r timm.layersr rrrrrrrr_builderr_efficientnet_blocksrr_efficientnet_builderrrrrrr _featuresr _features_fxr! _manipulater"r# _registryr$r%__all__rr)r&r'rfrgrrrerr+ default_cfgsr-r.rrCrBrGsCC E+J&&+13< . /OOOdX"))Xve(e(P#4N`. # 4 FX  %(  pp!p p  ppfc%    #'#& $ T @R  K KrC