o wiT@sdZddlZddlmZmZddlZddlZddlmZddlm Z m Z m Z ddl m Z ddlmZmZmZdd lmZdd lmZmZd d lmZeeZd edefddZd/deeefdefddZ Gdddej!Z"Gdddej#Z$Gdddej!Z%Gdddej!Z&Gdddej!Z'Gd d!d!ej!Z(Gd"d#d#ej!Z)Gd$d%d%ej!Z*eGd&d'd'eZ+eGd(d)d)e+Z,ed*d+Gd,d-d-e+Z-gd.Z.dS)0zPyTorch EfficientNet model.N)OptionalUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss)ACT2FN)BaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)EfficientNetConfigconfig num_channelscCsJ|j}||j9}t|t||d||}|d|kr!||7}t|S)z< Round number of filters based on depth multiplier. g?) depth_divisorwidth_coefficientmaxint)rrdivisornew_dimrs/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/efficientnet/modeling_efficientnet.py round_filters's   rT kernel_sizeadjustcCsnt|tr ||f}|dd|ddf}|r)|dd|d|dd|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)rrcorrectrrr correct_pad6s $r"cs<eZdZdZdeffdd ZdejdejfddZZ S) EfficientNetEmbeddingszL A module that corresponds to the stem module of the original work. rcshtt|d|_tjdd|_tj|j|jddddd|_ tj |j|j |j d |_ t|j|_dS) N )rrrrpaddingrrvalidFrstrider&bias)epsmomentum)super__init__rout_dimr ZeroPad2dr&Conv2dr convolution BatchNorm2dbatch_norm_epsbatch_norm_momentum batchnormr hidden_act activationselfr __class__rrr.Os  zEfficientNetEmbeddings.__init__ pixel_valuesreturncCs,||}||}||}||}|SN)r&r2r6r8)r:r=featuresrrrforwardZs    zEfficientNetEmbeddings.forward) __name__ __module__ __qualname____doc__rr.torchTensorrA __classcell__rrr;rr#Js r#cs,eZdZ       dfdd ZZS) EfficientNetDepthwiseConv2drrrTzerosc s*||} tj|| |||||||d dS)N) in_channels out_channelsrr)r&dilationgroupsr* padding_mode)r-r.) r:rKdepth_multiplierrr)r&rMr*rOrLr;rrr.ds  z$EfficientNetDepthwiseConv2d.__init__)rrrrrTrJ)rBrCrDr.rHrrr;rrIcsrIcsHeZdZdZdedededeffdd Zdejd ej fd d Z Z S) EfficientNetExpansionLayerz_ This corresponds to the expansion phase of each block in the original implementation. rin_dimr/r)csBttj||dddd|_tj||jd|_t|j |_ dS)NrsameFrKrLrr&r*) num_featuresr+) r-r.rr1 expand_convr3r4 expand_bnr r7 expand_act)r:rrRr/r)r;rrr.s z#EfficientNetExpansionLayer.__init__ hidden_statesr>cCs"||}||}||}|Sr?)rVrWrXr:rYrrrrAs   z"EfficientNetExpansionLayer.forward) rBrCrDrErrr.rF FloatTensorrGrArHrrr;rrQ}s rQc sLeZdZdZdededededef fdd Zd ej d ej fd d Z Z S) EfficientNetDepthwiseLayerzk This corresponds to the depthwise convolution phase of each block in the original implementation. rrRr)radjust_paddingcsvt||_|jdkrdnd}t||d}tj|d|_t||||dd|_tj ||j |j d|_ t |j|_dS) Nrr'rS)rr%Fr(rUr+r,)r-r.r)r"rr0depthwise_conv_padrIdepthwise_convr3r4r5depthwise_normr r7 depthwise_act)r:rrRr)rr]conv_padr&r;rrr.s    z#EfficientNetDepthwiseLayer.__init__rYr>cCs6|jdkr ||}||}||}||}|S)Nr)r)r_r`rarbrZrrrrAs     z"EfficientNetDepthwiseLayer.forward rBrCrDrErrboolr.rFr[rGrArHrrr;rr\sr\c sJeZdZdZd dedededeffdd Zd ej d ej fd d Z Z S)EfficientNetSqueezeExciteLayerzl This corresponds to the Squeeze and Excitement phase of each block in the original implementation. FrrR expand_dimexpandcst|r |n||_tdt||j|_tjdd|_ tj |j|jddd|_ tj |j|jddd|_ t |j|_t|_dS)Nr) output_sizerS)rKrLrr&)r-r.dimrrsqueeze_expansion_ratiodim_serAdaptiveAvgPool2dsqueezer1reducerhr r7 act_reduceSigmoid act_expand)r:rrRrgrhr;rrr.s$  z'EfficientNetSqueezeExciteLayer.__init__rYr>cCsF|}||}||}||}||}||}t||}|Sr?)rnrorprhrrrFmul)r:rYinputsrrrrAs      z&EfficientNetSqueezeExciteLayer.forward)Frdrrr;rrfs rfc sVeZdZdZdedededededef fdd Zd e j d e j d e j fd dZ Z S)EfficientNetFinalBlockLayerz[ This corresponds to the final phase of each block in the original implementation. rrRr/r) drop_rateid_skipcsXt|dko | |_tj||dddd|_tj||j|jd|_ tj |d|_ dS)NrrSFrTr^p) r-r. apply_dropoutrr1 project_convr3r4r5 project_bnDropoutdropout)r:rrRr/r)rvrwr;rrr.s  z$EfficientNetFinalBlockLayer.__init__ embeddingsrYr>cCs0||}||}|jr||}||}|Sr?)r{r|rzr~)r:rrYrrrrAs   z#EfficientNetFinalBlockLayer.forwardrBrCrDrErrfloatrer.rFr[rGrArHrrr;rrus $rucs\eZdZdZdededededededed ed effd d Zd e j de j fddZ Z 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. rrRr/r) expand_ratiorrvrwr]c st||_|jdkrdnd|_||} |jr"t||| |d|_t||jr)| n|||| d|_t||| |jd|_ t ||jrB| n|||||d|_ dS)NrTF)rrRr/r))rrRr)rr])rrRrgrh)rrRr/r)rvrw) r-r.rrhrQ expansionr\r`rfsqueeze_exciteru projection) r:rrRr/r)rrrvrwr] expand_in_dimr;rrr.!s4     zEfficientNetBlock.__init__rYr>cCs<|}|jdkr ||}||}||}|||}|S)Nr)rrr`rr)r:rYrrrrrAJs     zEfficientNetBlock.forwardrrrr;rrs,  )rc sPeZdZdZdeffdd Z  d dejdee d ee d e fd d Z Z S)EfficientNetEncoderz Forward propagates the embeddings through each EfficientNet block. Args: config ([`EfficientNetConfig`]): Model configuration class. rc s~t|_|j_fddt|j}tfdd|jD}d}g}t|D]k}t ||j|}t ||j |}|j |} |j |} |j |} t|j|D]@} | dkr]dnd} | dkredn| } | dkrm|n|}||jvrvdnd}|j||}t|||| | | || |d }|||d7}qUq+t|_tj|t |d dd dd _tj|j|j|jd _t|j_dS)Ncsttj|Sr?)rmathceildepth_coefficient)repeats)r:rr round_repeatsesz3EfficientNetEncoder.__init__..round_repeatsc3s|]}|VqdSr?r).0n)rrr jsz/EfficientNetEncoder.__init__..rTFr) rrRr/r)rrrvrwr]irSrTr^)r-r.rrlenrKsumnum_block_repeatsrangerrLstrides kernel_sizes expand_ratiosdepthwise_paddingdrop_connect_raterappendr ModuleListblocksr1top_convr3 hidden_dimr4r5top_bnr r7top_activation)r:rnum_base_blocks num_blockscurr_block_numrirRr/r)rrjrwr]rvblockr;)rr:rr.`s\          zEfficientNetEncoder.__init__FTrYoutput_hidden_states return_dictr>cCst|r|fnd}|jD] }||}|r||f7}q ||}||}||}|s4tdd||fDSt||dS)Ncss|] }|dur|VqdSr?r)rvrrrrsz.EfficientNetEncoder.forward..)last_hidden_staterY)rrrrtupler )r:rYrrall_hidden_statesrrrrrAs     zEfficientNetEncoder.forward)FT) rBrCrDrErr.rFr[rrer rArHrrr;rrWs:rc@s$eZdZeZdZdZgZddZdS)EfficientNetPreTrainedModel efficientnetr=cCstt|tjtjfr#|jjjd|jjd|j dur!|j j dSdSt|tj r8|j j |jj ddSdS)zInitialize the weightsg)meanstdNg?) r rLinearr1weightdatanormal_rinitializer_ranger*zero_ LayerNormfill_)r:modulerrr _init_weightss   z)EfficientNetPreTrainedModel._init_weightsN) rBrCrDr config_classbase_model_prefixmain_input_name_no_split_modulesrrrrrrs  rc s^eZdZdeffdd Ze   d deejdee dee de e e ffd d Z ZS) EfficientNetModelrcs~t|||_t||_t||_|jdkr"tj |j dd|_ n|jdkr1tj |j dd|_ nt d|j|dS)NrT) ceil_moderz2config.pooling must be one of ['mean', 'max'] got )r-r.rr#rrencoder pooling_typer AvgPool2drpooler MaxPool2d ValueErrorpooling post_initr9r;rrr.s      zEfficientNetModel.__init__Nr=rrr>cCs|dur|n|jj}|dur|n|jj}|durtd||}|j|||d}|d}||}||jdd}|sH||f|ddSt |||j dS)Nz You have to specify pixel_valuesrrrrr)r pooler_outputrY) rruse_return_dictrrrrreshapeshaper rY)r:r=rrembedding_outputencoder_outputsr pooled_outputrrrrAs*  zEfficientNetModel.forward)NNN)rBrCrDrr.rrrFr[rerrr rArHrrr;rrs rz EfficientNet Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for ImageNet. ) custom_introcsdeZdZfddZe    d deejdeejdee dee de e e ff d d Z ZS) "EfficientNetForImageClassificationcsdt||j|_||_t||_tj|jd|_ |jdkr't |j |jnt |_ |dS)Nrxr)r-r. num_labelsrrrrr} dropout_rater~rrIdentity classifierrr9r;rrr.s  $ z+EfficientNetForImageClassification.__init__Nr=labelsrrr>c Csl|dur|n|jj}|j|||d}|r|jn|d}||}||}d}|dur|jjdurU|jdkr;d|j_n|jdkrQ|jt j ksL|jt j krQd|j_nd|j_|jjdkrst } |jdkrm| | | }n+| ||}n%|jjdkrt} | |d|j|d}n|jjdkrt} | ||}|s|f|dd} |dur|f| S| St|||jd 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)losslogitsrY)rrrrr~r problem_typerdtyperFlongrrrnrviewrr rY) r:r=rrroutputsrrrloss_fctoutputrrrrAs@     "       z*EfficientNetForImageClassification.forward)NNNN)rBrCrDr.rrrFr[ LongTensorrerrr rArHrrr;rrs$   r)rrr)T)/rErtypingrrrFtorch.utils.checkpointrtorch.nnrrr activationsr modeling_outputsr r r modeling_utilsr utilsrrconfiguration_efficientnetr get_loggerrBloggerrrrrer"Moduler#r1rIrQr\rfrurrrrr__all__rrrrs@     ''!QZ8 E