^j ddlmZddlmZmZmZmZmZmZm Z ddl Z ddl m Z ddl m cmZddlmZmZddlmZmZmZmZmZmZmZmZmZddlmZddlm Z m!Z!dd l"m#Z#m$Z$m%Z%m&Z&m'Z'dd l(m)Z)dd l*m+Z+dd l,m-Z-dd l.m/Z/m0Z0ddgZ1ee jddZ3e+Gdde jhZ5Gdde jhZ6Gdde jhZ7dee8e jrfdee8e jrffdZ:d/de8de;de7fdZde;de;de7f dZ?d1d e8fd!Z@e/e@d"d#d$d%e@d&d"d#d$dd'(e@d)*d+ZAe0d/de;de7fd,ZBe0d/de;de6fd-ZCe0d/de;de6fd.ZDy)2)partial)CallableDictListOptionalSequenceTupleUnionNIMAGENET_INCEPTION_MEANIMAGENET_INCEPTION_STD) SelectAdaptivePool2dLinear LayerType RmsNorm2d ConvNormAct create_conv2dget_norm_layerget_norm_act_layer to_2tuple)build_model_with_cfg) SqueezeExciteUniversalInvertedResidual) BlockArgsEfficientNetBuilderdecode_arch_defefficientnet_init_weightsround_channels)feature_take_indices)register_notrace_module)checkpoint_seq)generate_default_cfgsregister_model MobileNetV5MobileNetV5Encodertanh) approximateceZdZdZ 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 | | d} t |t|tr t |n||_||_t||_||_ ||_ ||_ ||_ |xst}| xst} td|j |j d|j|| |j|jd| |_| |j fi| |_y)Ndevicedtyper)r+r,dw_kernel_size_mid exp_ratior2r3r1r0)super__init__ isinstancersum in_channels out_channelsrr-r.r/r0r1_GELUrrffnnorm)selfr+r,r-r.r/r0r1r2r3r6r7dd __class__s b/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/timm/models/mobilenetv5.pyr<z+MobileNetV5MultiScaleFusionAdapter.__init__4sU +B G&0&Bs6{DD&'89D*D0D"8DDK"UI(yJ( !!&&{{#::   DH4,,33DIinputsreturnc|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/appendtorchcatrBr- avg_pool2drC) rDrIhigh_resolutionresized_inputs_img feat_sizechannel_cat_imgs h_strides w_stridess rGforwardz*MobileNetV5MultiScaleFusionAdapter.forward\sQioobc*ONF##3IIbcN Qd,eee efd3ed4ed1efd5Zd+ej2d(ej2fd6Zd:d+ej2d7ed(ej2fd8Zd+ej2d(ej2fd9Z xZ!S)?r%z MobiletNet-V5 TFrLNavg block_args num_classesin_chans stem_size stem_biasfix_stem num_featurespad_typeuse_msfa msfa_indicesmsfa_output_resolutionr2r3aa_layerse_layer se_from_exp round_chs_fn drop_ratedrop_path_rater0 global_poolct|||d}| xst} t| xst} t | | }|xst }||_||_||_ d|_ | |_ | |_ |s||}t||fdd||| | d||_tdd|||| | ||||d |}t!j"||||_|j&|_|j(Dcgc]}|d c}|_|j,|_| r|x|_|_t3t5|j(|jd |_ t7|jDcgc]}|j(|d c}|_t;d|j8||j| | d ||_t?| |_ d |_!d |_"n|j,|_||_d |_t?| |_ |j.|j@jGz}tI||j0dfd|i||_!||j0fi||_"|rt!jJdnt!jL|_'|d kDrtQ|j0|fi|nt!jL|_)tU|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. r5FrrrQrRpaddingbiasr3r2 output_striderrrr2r3rrrr0stagernum_chsr+r,r-r3r2 pool_typeNrrr:)+r;r<rArrrrr{r|rgrad_checkpointingrrr conv_stemrnn Sequentialblocksfeatures feature_info stage_endsr+rhead_hidden_sizer lenr> msfa_in_chsr*msfarr conv_head norm_head feat_multrFlattenIdentityflattenr classifierr)rDrzr{r|r}r~rrrrrrr2r3rrrrrrr0rr6r7rEnorm_act_layerbuilderfminum_pooled_chsrFs rGr<zMobileNetV5.__init__s^  /& #J/<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;PRSl]eliklDN+D,A,AHRHDN(3rzz!} NY\]o&!6!6 JrJcecncncp!$'EB$bs ( K!Kc|j|jg}|j|j|j |j |j |j |j |j|j |j|jtjtj|j|jgtj|SN)rbn1extendrrWrrrrrDropoutrrr)rDlayerss rG as_sequentialzMobileNetV5.as_sequentials..$((+ dkk" d&&' >> % MM$.. ) >> % MM$.. ) rzz|RZZ%?QR}}f%%rHcoarsec.td|rdSdS)Nz^conv_stem|bn1z^blocks\.(\d+)z^blocks\.(\d+)\.(\d+))stemr)dict)rDrs rG group_matcherzMobileNetV5.group_matcher s%"(.$  4L  rHenablec||_yr)r)rDrs rGset_grad_checkpointingz"MobileNetV5.set_grad_checkpointings "(rHrJc|jSr)r)rDs rGget_classifierzMobileNetV5.get_classifiers rHc||_t||_|rtjdntj |_|dkDrt|j||_ ytj |_ y)Nrrr) r{rrrrrrrrr)rDr{rs rGreset_classifierzMobileNetV5.reset_classifiers[&/+F(3rzz!} HSVW&!6!6 D]_]h]h]jrHxindicesrC 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 rrrrrWrXjit is_scripting)rDrrrCrrrr intermediates take_indices max_indexifeat_idxrblks rGforward_intermediatesz!MobileNetV5.forward_intermediates"sB.Y&D(DD& % X'X X% &:3t{{;Ka;OQX&Y #L)&:3t;OQX&Y #L)8DE1DOOA.ELE 2I NN1  | #   # 99 ! ! #:[[F[[),F (C MHAA<'$$Q'  (  -;Fs'D$ prune_norm prune_headcp|r&tt|jdz|\}}n1tt|j|\}}|j|}|jd||_|t|jkrd|_d|_|r d|_d|_|j dd|S)z@ Prune layers not required for specified intermediates. rNrru)r rrrrrr)rDrrrrrrs rGprune_intermediate_layersz%MobileNetV5.prune_intermediate_layers`s &:3t{{;Ka;OQX&Y #L)&:3t;OQX&Y #L) 2Ikk*9- s4;;' '!DN!DN !DN!DN  ! !!R (rHc|jd}g}|j|}||jvr|j||jD]/}|dz }||}||jvs|j|1|j|}|S|j|}|j r8t jjst|j|d}|S|j |}|S)NrrT)r) rrrrWrrrXrrr"rDrrrrs rGforward_featureszMobileNetV5.forward_featuresxs 99 HMq!A4,,,$$Q'{{ ,A Ft000!((+  ,  -(A q!A&&uyy/E/E/G"4;;4@KKNrH pre_logitscZ|j|}|j|j|}|j|j|}|j|}|jdkDr,t j ||j|j}|r|S|j|S)Nrx)ptraining) rrrrrrUdropoutrr)rDrrs rG forward_headzMobileNetV5.forward_heads   Q  >> %q!A >> %q!A LLO >>B  !t~~ FA Hq!!rHcJ|j|}|j|}|Sr)rrrDrs rGrczMobileNetV5.forwards'  ! !! $   a rHF)T)ryNFFrFF)rFTF)"rerfrgrhrrrirlrkrrrrrjr<rrXrignorerrrModulerrrmr r rrrrrcrnros@rGr%r%~s( $"" $!&.*,-1.2,0,0 $%3!$&6:$1}(!}(}( }(  }(  }(}(}(}(}(s)}(%(}( *}(!+}(y)}( y)!}("#}($#%}(&'}((")}(*%-UO+}(,-}(~ & YY D   YY)T)) YY kCkck8<$$',!&< ||< eCcN34<  <  <  < !%< <  tELL!5tELL7I)I#JJ K< @./$#!& 3S >*   0%,,5<<. "ell " " "%,,rHc$eZdZdZdddddddd d d d ded d d d d fd ed ed edededede edede e de e de e de e dede de de de e f"fd Z d)dej de eeeefded ed!ed"ed#ed$eeej eej eej fffd%Zdej d$ej fd&Zdej d$ej fd'Zdej d$ej fd(ZxZS)*r&zMobileNetV5 Vision Encoderrr@TFrurvrsNrxrzr|r}r~rrrrr2r3rrrrrrr0c@t|||d}| xst} t| xst} | xst } d|_||_||_d|_ |s||}t||fdd||| | d||_ td d||| | | | | ||d |}tj||||_|j |_|j"Dcgc]}|d  c}|_d x|_|_t+t-|j"|d|_t1|j.Dcgc]}|j"|d c}|_||_t7d |j2|j&|j4| | d ||_t;|ycc}wcc}w)Nr5rFrrrrrrrrtrrr:)r;r<rArrrr{r|rrrrrrrrrrrrrr rrr>rrr*rr)rDrzr|r}r~rrrrr2r3rrrrrrr0r6r7rErrrrFs rGr<zMobileNetV5Encoder.__init__s,  /& #J/<9 ,}  ""'$Y/I$    !    &  %#!)#9    mmWY %CD #,,/3/@/@A!1W:A488D10T5F5F1GVWXY4K\K\]R 1 1" 5i @]^&<#6 ##%%"99!      "$'#B ^s  F*FrrrCrrrrrJc~|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 rrrrrWrrXrrr)rDrrrCrrrrrmsfa_intermediatesrrrrrrs rGrz(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<'$$Q'4,,,"))!,  -  yy+,m;;7Fs*E4cd}g}|j|}||jvr|j||jD]/}|dz }||}||jvs|j|1|j |S)Nrr)rrrWrrrs rGrz#MobileNetV5Encoder.forward_features;s NN1  t(( (   #;; (C MHAA4,,,$$Q'  (yy''rHctd)Nz=MobileNetV5Encoder does not support classification use cases.)NotImplementedErrorrs rGrzMobileNetV5Encoder.forward_headLs!"abbrHc$|j|Sr)rrs rGrczMobileNetV5Encoder.forwardOs$$Q''rHr)rerfrgrhrrrirlrkrrrrrjr<rXrmr rr rrrrcrnros@rGr&r&s$ ""*2*,-1.2,0,0 $%3!$&6:)O(!O(O( O(  O(  O(O(#3-O(%(O( *O(!+O(y)O(y)O(O(#O( !O(""#O($%-UO%O(h8<$$',!&C<||C<eCcN34C< C<  C<  C<!%C<C< tELL!5tELL7I)I#JJ KC)rzr}rrr3r2r0)rrrrrrArr) rrr,rrrarch_def model_kwargsrs rG_gen_mobilenet_v5r1sW$    & G4% p   & P( qa% F"8 C#c)^8JK#L //L$WjILI LWjALA LrHurlc 0|dddddttddd |S) Nrq)rrr4)rsrsr-bicubiczconv_stem.convr) r2r{ input_size pool_sizecrop_pct interpolationmeanstd first_convrr )r2rs rG_cfgr=9s04}S[)'0F&l     rH)rxrxrx)r-r-r-)rrr>)r:r;r6r{ztimm/gemma) hf_hub_idr:r;r6r{licenserq)r{)mobilenetv5_300m_enczmobilenetv5_300m.gemma3nzmobilenetv5_base.untrainedc J|jdd}t d|d|d|}|S)zMobileNet V5 Vision EncoderrsameT)rrr)rB)rr1)rrrrs rGrBrBYs@zz*f-H       E LrHc tdd|i|}|S)Nr)mobilenetv5_300mr1rrrs rGrFrFg  RZ R6 RE LrHc tdd|i|}|S)Nr)rrGrHs rGrrmrIrHr)r-NFF)ru)E functoolsrtypingrrrrrr r rXtorch.nnrtorch.nn.functional functionalrU timm.datar r timm.layersrrrrrrrrr_builderr_efficientnet_blocksrr_efficientnet_builderrrrrr _featuresr _features_fxr! _manipulater" _registryr#r$__all__GELUrArr*r%r&rkrmrrlrrrjr1r= default_cfgsrBrFrr:rHrGr\s"III E   +J,1'< . /V,SSSla"))aH n(n(b ell*+  #u||  #4N`0 # 4 k  %(  pp!p p  ppfc% |  !% |  !#'#!& , T @R  K KrH