rhzSdZddlZddlmZmZddlZddlZddlmZddlm Z m Z m Z ddl m Z ddlmZmZmZdd lmZdd lmZmZd d lmZej2eZd edefdZd,deeefdefdZ GddejBZ"GddejFZ$GddejBZ%GddejBZ&GddejBZ'GddejBZ(Gdd ejBZ)Gd!d"ejBZ*eGd#d$eZ+eGd%d&e+Z,ed'(Gd)d*e+Z-gd+Z.y)-zPyTorch EfficientNet model.N)OptionalUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss)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 /var/www/html/ai-insurance-compliance-backend/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)r r!corrects r correct_padr%6s}+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 )rrrrpaddingr rvalidFr strider+bias)epsmomentum)super__init__rout_dimr ZeroPad2dr+Conv2dr convolution BatchNorm2dbatch_norm_epsbatch_norm_momentum batchnormr hidden_act activationselfr __class__s rr3zEfficientNetEmbeddings.__init__Os $VR0 ||L9 99   1QPW^c  &:O:OZ`ZtZtu !2!23r pixel_valuesreturnc|j|}|j|}|j|}|j|}|SN)r+r7r;r=)r?rAfeaturess rforwardzEfficientNetEmbeddings.forwardZsA<< -##H->>(+??8,r) __name__ __module__ __qualname____doc__rr3torchTensorrF __classcell__r@s@rr'r'Js0 41 4ELLU\\rr'c.eZdZ dfd ZxZS)EfficientNetDepthwiseConv2dc @||z} t ||| ||||||| y)N) in_channels out_channelsr r.r+dilationgroupsr/ padding_mode)r2r3) r?rRdepth_multiplierr r.r+rTr/rVrSr@s rr3z$EfficientNetDepthwiseConv2d.__init__ds=#%55  #%#%  r)rr rrrTzeros)rGrHrIr3rMrNs@rrPrPcs$   rrPcleZdZdZdedededeffd Zdejdejfd Z xZ S) EfficientNetExpansionLayerz_ This corresponds to the expansion phase of each block in the original implementation. rin_dimr4r.ct|tj||ddd|_tj ||j |_t|j|_ y)NrsameFrRrSr r+r/) num_featuresr0) r2r3rr6 expand_convr8r9 expand_bnr r< expand_act)r?rr[r4r.r@s rr3z#EfficientNetExpansionLayer.__init__sZ 99   W&BWBWX !2!23r hidden_statesrBcl|j|}|j|}|j|}|SrD)r`rarbr?rcs rrFz"EfficientNetExpansionLayer.forwards4((7 }5  6 r) rGrHrIrJrrr3rK FloatTensorrLrFrMrNs@rrZrZ}sH 41 43 4 4VY 4U%6%65<<rrZc peZdZdZdededededef fd Zdejd ejfd Z xZ S) EfficientNetDepthwiseLayerzk This corresponds to the depthwise convolution phase of each block in the original implementation. rr[r.r adjust_paddingcbt|||_|jdk(rdnd}t||}t j ||_t||||d|_t j||j|j|_ t|j|_y) Nrr,r])r!r*Fr-r_r0r1)r2r3r.r%rr5depthwise_conv_padrPdepthwise_convr8r9r:depthwise_normr r< depthwise_act) r?rr[r.r riconv_padr+r@s rr3z#EfficientNetDepthwiseLayer.__init__s  "kkQ.7Fk.A"$,,w"?9  FHSX !nnV%:%:VE_E_ $F$5$56rrcrBc|jdk(r|j|}|j|}|j|}|j |}|S)Nr)r.rlrmrnrores rrFz"EfficientNetDepthwiseLayer.forwardsT ;;!  33MBM++M: ++M: **=9 r rGrHrIrJrrboolr3rKrfrLrFrMrNs@rrhrhsZ7"77 7  7  7, U%6%6 5<< rrhc neZdZdZd dedededeffd Zdejdejfd Z xZ S) EfficientNetSqueezeExciteLayerzl This corresponds to the Squeeze and Excitement phase of each block in the original implementation. rr[ expand_dimexpandct||r|n||_tdt ||j z|_tjd|_ tj|j|j dd|_ tj|j |jdd|_ t|j|_tj |_y)Nr) output_sizer])rRrSr r+)r2r3dimrrsqueeze_expansion_ratiodim_serAdaptiveAvgPool2dsqueezer6reducerwr r< act_reduceSigmoid act_expand)r?rr[rvrwr@s rr3z'EfficientNetSqueezeExciteLayer.__init__s !':V!S&*H*H!HIJ ++: ii   ii    !!2!23**,rrcrBc|}|j|}|j|}|j|}|j|}|j |}t j ||}|SrD)r~rrrwrrKmul)r?rcinputss rrFz&EfficientNetSqueezeExciteLayer.forwardsc ]3  M2  6  M2  6  &-8 r)FrrrNs@rrurusH'1'3'C'Y]'* U%6%6 5<< rruc eZdZdZdedededededef fd Zd e jd e jd e jfd Z xZ S) EfficientNetFinalBlockLayerz[ This corresponds to the final phase of each block in the original implementation. rr[r4r. drop_rateid_skipct||dk(xr| |_tj||ddd|_tj ||j|j|_ tj||_ y)Nrr]Fr^rkp) r2r3 apply_dropoutrr6 project_convr8r9r: project_bnDropoutdropout)r?rr[r4r.rrr@s rr3z$EfficientNetFinalBlockLayer.__init__sz #q[8[II   .. f&;&;fF`F` zzI. r embeddingsrcrBc|j|}|j|}|jr|j|}||z}|SrD)rrrr)r?rrcs rrFz#EfficientNetFinalBlockLayer.forwardsG))-8  6    LL7M)J6MrrGrHrIrJrrfloatrsr3rKrfrLrFrMrNs@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. rr[r4r. expand_ratior rrric lt |||_|jdk7|_||z} |jrt ||| ||_t ||jr| n|||| |_t||| |j|_ t||jr| n||||||_ y)Nr)rr[r4r.)rr[r.r ri)rr[rvrw)rr[r4r.rr) r2r3rrwrZ expansionrhrmrusqueeze_exciter projection) r?rr[r4r.rr rrri expand_in_dimr@s rr3zEfficientNetBlock.__init__!s (''1, - ;;7fmFDN9$(KK=V#)  =&]4;; 6$(KK=V  rrcrBc|}|jdk7r|j|}|j|}|j|}|j ||}|S)Nr)rrrmrr)r?rcrs rrFzEfficientNetBlock.forwardJs["    ! NN=9M++M: ++M:  MB rrrNs@rrrs4' "' '  '  '  ' ' ' ' ' R U%6%6 5<< rrc feZdZdZdeffd Z d dejdee dee de fdZ 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|zSrD)rmathceildepth_coefficient)repeatsr?s r round_repeatsz3EfficientNetEncoder.__init__..round_repeatses"tyy!7!7'!ABC Crc3.K|] }|ywrD).0nrs r z/EfficientNetEncoder.__init__..jsLaq)Lsrr) rr[r4r.r rrrriir]Fr^rk)r2r3rrlenrRsumnum_block_repeatsrangerrSstrides kernel_sizes expand_ratiosdepthwise_paddingdrop_connect_raterappendr ModuleListblocksr6top_convr8 hidden_dimr9r:top_bnr r<top_activation)r?rnum_base_blocks num_blockscurr_block_numrir[r4r.r rjrrirblockrr@s` @rr3zEfficientNetEncoder.__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$!#' $ $8mmF+  &vt4   nn**0E0EPVPjPj  %V%6%67rrcoutput_hidden_states return_dictrBc|r|fnd}|jD]}||}|s||fz }|j|}|j|}|j|}|st d||fDSt ||S)Nc3&K|] }|| ywrDr)rvs rrz.EfficientNetEncoder.forward..sXq!-Xs)last_hidden_staterc)rrrrtupler )r?rcrrall_hidden_statesrs rrFzEfficientNetEncoder.forwards 1E],$[[ 6E!-0M#!m%55! 6  m4  M2 ++M: X]4E$FXX X-++  r)FT) rGrHrIrJrr3rKrfrrsr rFrMrNs@rrrWsW58158t05&*  (( 'tn d^  (  rrcDeZdZUeed<dZdZgZdejfdZ y)EfficientNetPreTrainedModelr efficientnetrAmodulecPt|tjtjtjfrm|j j jd|jj|j%|jj jyyy)zInitialize the weightsg)meanstdN) r#rLinearr6r8weightdatanormal_rinitializer_ranger/zero_)r?rs r _init_weightsz)EfficientNetPreTrainedModel._init_weightsss fryy"))R^^D E MM   & &CT[[5R5R & S{{&   &&(' FrN) rGrHrIr__annotations__base_model_prefixmain_input_name_no_split_modulesrModulerrrrrrs) &$O)BII)rrc ~eZdZdeffd Ze ddeejdee dee de e e ffdZ xZS) 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 )r2r3rr'rrencoder pooling_typer AvgPool2drpooler MaxPool2d ValueErrorpooling post_initr>s rr3zEfficientNetModel.__init__s   08*62    & (,,v'8'8DIDK  E ),,v'8'8DIDKQRXR`R`Qabc c rrArrrBc~||n|jj}||n|jj}| td|j |}|j |||}|d}|j |}|j|jdd}|s ||f|ddzSt|||jS)Nz You have to specify pixel_valuesrrrrr)r pooler_outputrc) rruse_return_dictrrrrreshapeshaper rc)r?rArrembedding_outputencoder_outputsr pooled_outputs rrFzEfficientNetModel.forwards%9$D $++JjJj &1%> ? " " rrz EfficientNet Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for ImageNet. ) custom_introceZdZfdZe ddeejdeejdee dee de e e ff dZ xZS) "EfficientNetForImageClassificationct|||j|_||_t ||_t j|j|_ |jdkDr*t j|j|jnt j|_ |jy)Nrr)r2r3 num_labelsrrrrr dropout_raterrrIdentity classifierrr>s rr3z+EfficientNetForImageClassification.__init__s   ++ -f5zzF$7$78 KO??]^K^"))F$5$5tGdfdododq rrAlabelsrrrBc:||n|jj}|j|||}|r |jn|d}|j |}|j |}d}||jj |jdk(rd|j_nl|jdkDrL|jtjk(s|jtjk(rd|j_nd|j_|jj dk(rIt} |jdk(r&| |j|j}n| ||}n|jj dk(r=t} | |jd|j|jd}n,|jj dk(rt!} | ||}|s|f|ddz} ||f| zS| St#|||j$ S) 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 regressionsingle_label_classificationmulti_label_classificationr)losslogitsrc)rrrrrr problem_typerdtyperKlongrrr~rviewrr rc) r?rArrroutputsrrrloss_fctoutputs rrFz*EfficientNetForImageClassification.forwards&1%rsq" " AA! .,:   H % , C @U3:.@@(RYY2 ")) 44$$N$RYY$N"))BN NbW "))W t )/ ) ) 5 35 5 p B )DB  B J er