sgEdZddlZddlmZddlZddlZddlmZmZddlm Z m Z m Z ddl m Z ddlmZmZmZdd lmZmZmZdd lmZdd lmZd d lmZej8eZdZdZ gdZ!dZ"dZ#GddejHZ%GddejHZ&GddejHZ'GddejHZ(GddejHZ)GddejHZ*GddejHZ+Gd d!ejHZ,Gd"d#eZ-d$Z.d%Z/ed&e.Gd'd(e-Z0ed)e.Gd*d+e-Z1y),zPyTorch RegNet model.N)Optional)Tensornn)BCEWithLogitsLossCrossEntropyLossMSELoss)ACT2FN)add_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forward)BaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)logging) RegNetConfigrzfacebook/regnet-y-040)ri@rztabby, tabby catcNeZdZ d dedededededeef fd ZdZxZS) RegNetConvLayer in_channels out_channels kernel_sizestridegroups activationc t|tj|||||dz|d|_tj ||_|t||_ ytj|_ y)NF)rrpaddingrbias) super__init__rConv2d convolution BatchNorm2d normalizationr Identityr)selfrrrrrr __class__s ]/var/www/html/venv/lib/python3.12/site-packages/transformers/models/regnet/modeling_regnet.pyr#zRegNetConvLayer.__init__4sk 99  #1$  ^^L90:0F&,BKKMcl|j|}|j|}|j|}|SN)r%r'rr) hidden_states r+forwardzRegNetConvLayer.forwardJs6'' 5 )),7 |4 r,)r rrrelu) __name__ __module__ __qualname__intrstrr#r1 __classcell__r*s@r+rr3s` $*ZZZ Z  Z  ZSMZ,r,rc.eZdZdZdeffd ZdZxZS)RegNetEmbeddingszP RegNet Embedddings (stem) composed of a single aggressive convolution. configct|t|j|jdd|j |_|j|_y)Nr r)rrr)r"r#r num_channelsembedding_size hidden_actembedderr)r<r*s r+r#zRegNetEmbeddings.__init__VsH '   !6!6Aa\b\m\m  #//r,cz|jd}||jk7r td|j|}|S)NrzeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)shaper> ValueErrorrA)r) pixel_valuesr>r0s r+r1zRegNetEmbeddings.forward]sF#))!, 4,, ,w }}\2 r,)r3r4r5__doc__rr#r1r8r9s@r+r;r;Qs0|0r,r;cBeZdZdZd dededeffd ZdedefdZxZS) RegNetShortCutz RegNet shortcut, used to project the residual features to the correct size. If needed, it is also used to downsample the input using `stride=2`. rrrct|tj||d|d|_tj ||_y)NrF)rrr!)r"r#rr$r%r&r')r)rrrr*s r+r#zRegNetShortCut.__init__ns: 99[,AV\chi^^L9r,inputreturncJ|j|}|j|}|Sr.)r%r')r)rKr0s r+r1zRegNetShortCut.forwardss(''. )),7 r,)r) r3r4r5rGr6r#rr1r8r9s@r+rIrIhs5 :C:s:C: Vr,rIc2eZdZdZdedeffd ZdZxZS) RegNetSELayerz| Squeeze and Excitation layer (SE) proposed in [Squeeze-and-Excitation Networks](https://arxiv.org/abs/1709.01507). rreduced_channelsc 0t|tjd|_tj tj ||dtjtj ||dtj|_ y)Nrrr)r) r"r#rAdaptiveAvgPool2dpooler Sequentialr$ReLUSigmoid attention)r)rrPr*s r+r#zRegNetSELayer.__init__~sd **62  IIk#3 C GGI II&  C JJL  r,cT|j|}|j|}||z}|Sr.)rTrX)r)r0pooledrXs r+r1zRegNetSELayer.forwards.\*NN6* #i/ r,)r3r4r5rGr6r#r1r8r9s@r+rOrOys   C  3  r,rOc <eZdZdZddedededeffd ZdZxZS) RegNetXLayerzt RegNet's layer composed by three `3x3` convolutions, same as a ResNet bottleneck layer with reduction = 1. r<rrrc t|||k7xs|dk7}td||jz}|rt |||nt j |_t jt||d|jt|||||jt||dd|_ t|j|_ y)Nrrrrrrr)r"r#max groups_widthrIrr(shortcutrUrr@layerr rr)r<rrrshould_apply_shortcutrr*s r+r#zRegNetXLayer.__init__s  +| ; Jv{Q (;(;;<H]N; V Dcecncncp ]] K1QWQbQb c L,vfagarar s L,ARV W  !!2!23r,cz|}|j|}|j|}||z }|j|}|Sr.rdrcrr)r0residuals r+r1zRegNetXLayer.forwardAzz,/ ==* |4 r,r r3r4r5rGrr6r#r1r8r9s@r+r\r\s/ 4| 4# 4S 4Z] 4r,r\c <eZdZdZddedededeffd ZdZxZS) RegNetYLayerzC RegNet's Y layer: an X layer with Squeeze and Excitation. r<rrrc t|||k7xs|dk7}td||jz}|rt |||nt j |_t jt||d|jt|||||jt|tt|dz t||dd|_t|j|_y)Nrr^r_r`)rP)r"r#rarbrIrr(rcrUrr@rOr6roundrdr rres r+r#zRegNetYLayer.__init__s  +| ; Jv{Q (;(;;<H]N; V Dcecncncp ]] K1QWQbQb c L,vfagarar s ,U;QR?=S9T U L,ARV W   !!2!23r,cz|}|j|}|j|}||z }|j|}|Sr.rhris r+r1zRegNetYLayer.forwardrkr,rlrmr9s@r+roros/ 4| 4# 4S 4Z] 4r,roc DeZdZdZ d dededededef fd ZdZxZS) RegNetStagez4 A RegNet stage composed by stacked layers. r<rrrdepthc t||jdk(rtnt}t j |||||gt|dz Dcgc] }||||c}|_ycc}w)Nxr^r) r"r# layer_typer\rorrUrangelayers) r)r<rrrrvrd_r*s r+r#zRegNetStage.__init__sw  & 1 1S 8 lmm     BGuqyAQ RAeFL,7 R   SsA3 c(|j|}|Sr.)r{r/s r+r1zRegNetStage.forwards{{<0 r,)rrrmr9s@r+rurusJ          .r,ruc @eZdZdeffd Z ddedededefdZxZ S) RegNetEncoderr<c t|tjg|_|jj t ||j|jd|jrdnd|jdt|j|jdd}t||jddD]0\\}}}|jj t ||||2y)Nrrr)rrv)rv) r"r#r ModuleListstagesappendrur? hidden_sizesdownsample_in_first_stagedepthszip)r)r<in_out_channelsrrrvr*s r+r#zRegNetEncoder.__init__s mmB'   %%##A&"<z(RegNetEncoder.forward.. sSqQ]Ss)last_hidden_state hidden_states)rtupler)r)r0rrr stage_modules r+r1zRegNetEncoder.forwardsq3  KK 6L# - ? ' 5L  6 )\O;MS\=$ASS S- \ijjr,)FT) r3r4r5rr#rboolrr1r8r9s@r+rrsB\|\$]ak"k:>kUYk 'kr,rc(eZdZdZeZdZdZdgZdZ y)RegNetPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. regnetrFrocJt|tjr-tjj |j ddyt|tj rtjj|j tjd|jxtjj|j \}}|dkDrdtj|z nd}tjj|j| |yyt|tjtjfrUtjj|j dtjj|jdyy)Nfan_outr2)mode nonlinearity)arr) isinstancerr$initkaiming_normal_weightLinearkaiming_uniform_mathsqrtr!_calculate_fan_in_and_fan_outuniform_r& GroupNorm constant_)r)modulefan_inr|bounds r+ _init_weightsz#RegNetPreTrainedModel._init_weightss fbii ( GG # #FMM PV # W  * GG $ $V]]diil $ C{{&GGAA&--P 17!DIIf--  ufe<' > ? GG  fmmQ / GG  fkk1 -@r,N) r3r4r5rGr config_classbase_model_prefixmain_input_name_no_split_modulesrrr,r+rrs'  L $O'( .r,raI This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matters related to general usage and behavior. Parameters: config ([`RegNetConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. aK Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`ConvNextImageProcessor.__call__`] for details. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple. zOThe bare RegNet model outputting raw features without any specific head on top.c |eZdZfdZeeeeee de  d de de e de e defdZxZS) RegNetModelct||||_t||_t ||_tjd|_ |jy)NrR) r"r#r<r;rArencoderrrSrT post_initrBs r+r#zRegNetModel.__init__KsK   (0 $V, **62  r,vision) checkpoint output_typermodalityexpected_outputrFrrrLc(||n|jj}||n|jj}|j|}|j |||}|d}|j |}|s ||f|ddzSt |||jS)Nrrrr)r pooler_outputr)r<ruse_return_dictrArrTrr)r)rFrrembedding_outputencoder_outputsr pooled_outputs r+r1zRegNetModel.forwardTs%9$D $++JjJj &1% 1` a classification loss is computed (Cross-Entropy). Nrr regressionsingle_label_classificationmulti_label_classificationrr)losslogitsr)r<rrrr problem_typerdtypetorchlongr6rsqueezerviewrrr) r)rFrrroutputsrrrloss_fctoutputs r+r1z$RegNetForImageClassification.forwards&&1%rsm  AA!qq ..   H %!.(21bii<ryy.RYY"BII02998299:"))B#kBII#kL.O.6  U + '+  + \ Cs#8CsCsr,