iOdZddlZddlZddlmZddlmZddlmZddl m Z m Z m Z ddl mZdd lmZmZd d lmZej(eZd ed efdZd+deezdefdZGddej8ZGddej<ZGddej8Z Gddej8Z!Gddej8Z"Gddej8Z#Gddej8Z$Gd d!ej8Z%eGd"d#eZ&eGd$d%e&Z'ed&'Gd(d)e&Z(gd*Z)y),zPyTorch EfficientNet model.N)nn)initialization)ACT2FN)BaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)EfficientNetConfigconfig num_channelsc|j}||jz}t|t||dz z|z|z}|d|zkr||z }t|S)z< Round number of filters based on depth multiplier. g?) depth_divisorwidth_coefficientmaxint)rrdivisornew_dims r/mnt/e/genesis-system/.venv/lib/python3.12/site-packages/transformers/models/efficientnet/modeling_efficientnet.py round_filtersr$sf""GF,,,L'3|gk9:gEOPG|##7 w< kernel_sizeadjustct|tr||f}|ddz|ddzf}|r|ddz |d|ddz |dfS|d|d|d|dfS)aJ Utility function to get the tuple padding value for the depthwise convolution. Args: kernel_size (`int` or `tuple`): Kernel size of the convolution layers. adjust (`bool`, *optional*, defaults to `True`): Adjusts padding value to apply to right and bottom sides of the input. rrr ) isinstancer)rrcorrects r correct_padr!3s}+s#"K0 1~"KNa$78G  Q GAJNGAJGG GAJ GAJ??rc`eZdZdZdeffd ZdejdejfdZxZ S)EfficientNetEmbeddingszL A module that corresponds to the stem module of the original work. rct|t|d|_t j d|_t j|j|jdddd|_ t j|j|j|j |_ t|j|_y) N )rr rr paddingrrvalidFrstrider'bias)epsmomentum)super__init__rout_dimr ZeroPad2dr'Conv2dr convolution BatchNorm2dbatch_norm_epsbatch_norm_momentum batchnormr hidden_act activationselfr __class__s rr/zEfficientNetEmbeddings.__init__Ls $VR0 ||L9 99   1QPW^c  &:O:OZ`ZtZtu !2!23r pixel_valuesreturnc|j|}|j|}|j|}|j|}|SN)r'r3r7r9)r;r=featuress rforwardzEfficientNetEmbeddings.forwardWsA<< -##H->>(+??8,r) __name__ __module__ __qualname____doc__rr/torchTensorrB __classcell__r<s@rr#r#Gs0 41 4ELLU\\rr#c.eZdZ dfd ZxZS)EfficientNetDepthwiseConv2dc @||z} t ||| ||||||| y)N) in_channels out_channelsrr*r'dilationgroupsr+ padding_mode)r.r/) r;rNdepth_multiplierrr*r'rPr+rRrOr<s rr/z$EfficientNetDepthwiseConv2d.__init__as=#%55  #%#%  r)r rr rr Tzeros)rCrDrEr/rIrJs@rrLrL`s$   rrLcleZdZdZdedededeffd Zdejdejfd Z xZ S) EfficientNetExpansionLayerz_ This corresponds to the expansion phase of each block in the original implementation. rin_dimr0r*ct|tj||ddd|_tj ||j |_t|j|_ y)Nr sameFrNrOrr'r+) num_featuresr,) r.r/rr2 expand_convr4r5 expand_bnrr8 expand_act)r;rrWr0r*r<s rr/z#EfficientNetExpansionLayer.__init__sZ 99   W&BWBWX !2!23r hidden_statesr>cl|j|}|j|}|j|}|Sr@)r\r]r^r;r_s rrBz"EfficientNetExpansionLayer.forwards4((7 }5  6 r) rCrDrErFrrr/rG FloatTensorrHrBrIrJs@rrVrVzsH 41 43 4 4VY 4U%6%65<<rrVc peZdZdZdededededef fd Zdejd ejfd Z xZ S) EfficientNetDepthwiseLayerzk This corresponds to the depthwise convolution phase of each block in the original implementation. rrWr*radjust_paddingcbt|||_|jdk(rdnd}t||}t j ||_t||||d|_t j||j|j|_ t|j|_y) Nrr(rY)rr&Fr)r[r,r-)r.r/r*r!rr1depthwise_conv_padrLdepthwise_convr4r5r6depthwise_normrr8 depthwise_act) r;rrWr*rreconv_padr'r<s rr/z#EfficientNetDepthwiseLayer.__init__s  "kkQ.7Fk.A"$,,w"?9  FHSX !nnV%:%:VE_E_ $F$5$56rr_r>c|jdk(r|j|}|j|}|j|}|j |}|S)Nr)r*rhrirjrkras rrBz"EfficientNetDepthwiseLayer.forwardsT ;;!  33MBM++M: ++M: **=9 r rCrDrErFrrboolr/rGrbrHrBrIrJs@rrdrdsZ7"77 7  7  7, U%6%6 5<< rrdc neZdZdZd dedededeffd Zdejdejfd Z xZ S) EfficientNetSqueezeExciteLayerzl This corresponds to the Squeeze and Excitement phase of each block in the original implementation. rrW expand_dimexpandct||r|n||_tdt ||j z|_tjd|_ tj|j|j dd|_ tj|j |jdd|_ t|j|_tj |_y)Nr ) output_sizerY)rNrOrr')r.r/dimrrsqueeze_expansion_ratiodim_serAdaptiveAvgPool2dsqueezer2reducersrr8 act_reduceSigmoid act_expand)r;rrWrrrsr<s rr/z'EfficientNetSqueezeExciteLayer.__init__s !':V!S&*H*H!HIJ ++: ii   ii    !!2!23**,rr_r>c|}|j|}|j|}|j|}|j|}|j |}t j ||}|Sr@)rzr{r|rsr~rGmul)r;r_inputss rrBz&EfficientNetSqueezeExciteLayer.forwardsc ]3  M2  6  M2  6  &-8 r)FrnrJs@rrqrqsH'1'3'C'Y]'* U%6%6 5<< rrqc eZdZdZdedededededef fd Zd e jd e jd e jfd Z xZ S) EfficientNetFinalBlockLayerz[ This corresponds to the final phase of each block in the original implementation. rrWr0r* drop_rateid_skipct||dk(xr| |_tj||ddd|_tj ||j|j|_ tj||_ y)Nr rYFrZrgp) r.r/ apply_dropoutrr2 project_convr4r5r6 project_bnDropoutdropout)r;rrWr0r*rrr<s rr/z$EfficientNetFinalBlockLayer.__init__sz #q[8[II   .. f&;&;fF`F` zzI. r embeddingsr_r>c|j|}|j|}|jr|j|}||z}|Sr@)rrrr)r;rr_s rrBz#EfficientNetFinalBlockLayer.forwardsG))-8  6    LL7M)J6MrrCrDrErFrrfloatror/rGrbrHrBrIrJs@rrrsj/(/25/@C/MP/]b/mq/"%"3"3EDUDUZ_ZfZfrrceZdZdZdededededededed ed effd Zd e jd e jfdZ xZ S)EfficientNetBlocka This corresponds to the expansion and depthwise convolution phase of each block in the original implementation. Args: config ([`EfficientNetConfig`]): Model configuration class. in_dim (`int`): Number of input channels. out_dim (`int`): Number of output channels. stride (`int`): Stride size to be used in convolution layers. expand_ratio (`int`): Expand ratio to set the output dimensions for the expansion and squeeze-excite layers. kernel_size (`int`): Kernel size for the depthwise convolution layer. drop_rate (`float`): Dropout rate to be used in the final phase of each block. id_skip (`bool`): Whether to apply dropout and sum the final hidden states with the input embeddings during the final phase of each block. Set to `True` for the first block of each stage. adjust_padding (`bool`): Whether to apply padding to only right and bottom side of the input kernel before the depthwise convolution operation, set to `True` for inputs with odd input sizes. rrWr0r* expand_ratiorrrrec lt |||_|jdk7|_||z} |jrt ||| ||_t ||jr| n|||| |_t||| |j|_ t||jr| n||||||_ y)Nr )rrWr0r*)rrWr*rre)rrWrrrs)rrWr0r*rr) r.r/rrsrV expansionrdrirqsqueeze_exciter projection) r;rrWr0r*rrrrre expand_in_dimr<s rr/zEfficientNetBlock.__init__s (''1, - ;;7fmFDN9$(KK=V#)  =&]4;; 6$(KK=V  rr_r>c|}|jdk7r|j|}|j|}|j|}|j ||}|S)Nr )rrrirr)r;r_rs rrBzEfficientNetBlock.forwardGs["    ! NN=9M++M: ++M:  MB rrrJs@rrrs4' "' '  '  '  ' ' ' ' ' R U%6%6 5<< rrc feZdZdZdeffd Z d dejdedzdedzde fd Z xZ S) EfficientNetEncoderz Forward propagates the embeddings through each EfficientNet block. Args: config ([`EfficientNetConfig`]): Model configuration class. rc t|_|j_fdt |j }t fd|jD}d}g}t|D]}t||j |}t||j|}|j|} |j|} |j|} t|j|D]c} | dk(} | dkDrdn| } | dkDr|n|}||jv}|j|z|z }t!|||| | | || | }|j#||dz }et%j&|_t%j*t|dddd _t%j.|j0|j2|j4 _t8|j:_y) NcZttjj|zSr@)rmathceildepth_coefficient)repeatsr;s r round_repeatsz3EfficientNetEncoder.__init__..round_repeatsbs"tyy!7!7'!ABC Crc3.K|] }|ywr@).0nrs r z/EfficientNetEncoder.__init__..gsLaq)Lsrr ) rrWr0r*rrrrreirYFrZrg)r.r/rrlenrNsumnum_block_repeatsrangerrOstrides kernel_sizes expand_ratiosdepthwise_paddingdrop_connect_raterappendr ModuleListblocksr2top_convr4 hidden_dimr5r6top_bnrr8top_activation)r;rnum_base_blocks num_blockscurr_block_numrirWr0r*rrjrrerblockrr<s` @rr/zEfficientNetEncoder.__init__]s  !'!9!9 Df001L63K3KLL ' $A"66+=+=a+@AF#FF,?,?,BCG^^A&F --a0K!//2L=)A)A!)DEF $q&!e$%Ev!/v7O7O!O"44~E R )!!#! +!-'##1  e$!#' $ $8mmF+  &vt4   nn**0E0EPVPjPj  %V%6%67rr_output_hidden_statesN return_dictr>c|r|fnd}|jD]}||}|s||fz }|j|}|j|}|j|}|st d||fDSt ||S)Nc3&K|] }|| ywr@r)rvs rrz.EfficientNetEncoder.forward..sXq!-Xs)last_hidden_stater_)rrrrtupler)r;r_rrall_hidden_statesrs rrBzEfficientNetEncoder.forwards 1E],$[[ 6E!-0M#!m%55! 6  m4  M2 ++M: X]4E$FXX X-++  r)FT) rCrDrErFrr/rGrbrorrBrIrJs@rrrTsW58158t-2#'  (( #Tk D[  (  rrcreZdZUeed<dZdZdZdgZe jde jfdZ y) EfficientNetPreTrainedModelr efficientnetr=)imagermodulect|tjtjtjfrt j |jd|jj|jt j|jt|dd^t j|jt j|jt j|j yyy)zInitialize the weightsg)meanstdN running_mean)rrLinearr2r4initnormal_weightrinitializer_ranger+zeros_getattrrones_ running_varnum_batches_tracked)r;rs r _init_weightsz)EfficientNetPreTrainedModel._init_weightss fryy"))R^^D E LLSdkk6S6S T{{& FKK(v~t4@ F//0 6--. F667A FrN)rCrDrEr__annotations__base_model_prefixmain_input_nameinput_modalities_no_split_modulesrGno_gradrModulerrrrrrsG &$O!,-U]]_ 8BII 8 8rrc zeZdZdeffd Ze d dejdzdedzdedzde e zfdZ xZ S) EfficientNetModelrct||||_t||_t ||_|jdk(r'tj|jd|_ nN|jdk(r'tj|jd|_ ntd|j|jy)NrT) ceil_moderz2config.pooling must be one of ['mean', 'max'] got )r.r/rr#rrencoder pooling_typer AvgPool2drpooler MaxPool2d ValueErrorpooling post_initr:s rr/zEfficientNetModel.__init__s   08*62    & (,,v'8'8DIDK  E ),,v'8'8DIDKQRXR`R`Qabc c rNr=rrr>c ~||n|jj}||n|jj}| td|j |}|j |||}|d}|j |}|j|jdd}|s ||f|ddzSt|||jS)Nz You have to specify pixel_valuesrrrrr )r pooler_outputr_) rruse_return_dictrrrrreshapeshaperr_) r;r=rrkwargsembedding_outputencoder_outputsr pooled_outputs rrBzEfficientNetModel.forwards%9$D $++JjJj &1%c d||n|jj}|j|||}|r |jn|d}|j |}|j |}d} ||j |||j} |s|f|ddz} | | f| zS| St| ||jS)a labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Nrr r)losslogitsr_) rrrrrr loss_functionr r_) r;r=rrrroutputsrrroutputs rrBz*EfficientNetForImageClassification.forwards&1%rs`" &! .,:   H % , C @S5[@$@(RYY2 ")) 44$$N$RYY$N"))BN NbW "))W t8/88(6 36 6 r 0 )D0  0 f er