rh@"dZddlmZmZmZddlZddlmZddlmZddlm Z m Z m Z ddl m Z ddlmZddlmZmZmZmZdd lmZdd lmZdd lmZd d lmZej:eZGddej@Z!Gddej@Z"Gddej@Z#Gddej@Z$eGdde Z%eGdde%Z&edGdde%Z'edGdd e%eZ(gd!Z)y)"zPyTorch TextNet model.)AnyOptionalUnionN)Tensor)BCEWithLogitsLossCrossEntropyLossMSELoss)PreTrainedModel)ACT2CLS)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention) TextNetConfig)logging) BackboneMixin)auto_docstringc\eZdZdeffd Zdej dej fdZxZS)TextNetConvLayerconfigct||j|_|j|_|j |_t|jtr$|jddz|jddzfn|jdz}tj|j|j|j|j|d|_tj|j|j |_tj$|_|jt)|j|_yy)NrF) kernel_sizestridepaddingbias)super__init__stem_kernel_sizer stem_strider stem_act_funcactivation_function isinstancetuplennConv2dstem_num_channelsstem_out_channelsconv BatchNorm2dbatch_norm_eps batch_normIdentity activationr )selfrr __class__s /var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/transformers/models/textnet/modeling_textnet.pyr zTextNetConvLayer.__init__+s !22(( #)#7#7 &1159   "a '););A)>!)C D((A-  II  $ $  $ $//%%   ..)A)A6CXCXY++-  # # /%d&>&>?ADO 0 hidden_statesreturnch|j|}|j|}|j|SN)r+r.r0)r1r5s r3forwardzTextNetConvLayer.forwardFs- -0  6 }--r4) __name__ __module__ __qualname__rr torchrr9 __classcell__r2s@r3rr*s,B}B6.U\\.ell.r4rc peZdZdZdededededef fd Zdejd ejfd Z xZ S) TextNetRepConvLayera This layer supports re-parameterization by combining multiple convolutional branches (e.g., main convolution, vertical, horizontal, and identity branches) during training. At inference time, these branches can be collapsed into a single convolution for efficiency, as per the re-parameterization paradigm. The "Rep" in the name stands for "re-parameterization" (introduced by RepVGG). r in_channels out_channelsrrctt |||_||_||_||_|ddz dz|ddz dzf}t j|_t j|||||d|_ t j||j|_ |ddz dzdf}d|ddz dzf}|ddk7rLt j|||ddf||d|_t j||j|_nd\|_|_|ddk7rLt j||d|df||d|_t j||j|_nd\|_|_||k(r,|dk(r't j||j|_yd|_y)NrrrF)rBrCrrrr) num_featuresepsNN)rr num_channelsrCrrr'ReLUr$r( main_convr,r-main_batch_norm vertical_convvertical_batch_normhorizontal_convhorizontal_batch_norm rbr_identity) r1rrBrCrrrvertical_paddinghorizontal_paddingr2s r3r zTextNetRepConvLayer.__init__Vs '(& NQ&1,{1~/Aa.GH#%779 #%#   "~~Q !#')(^Q/( "D (*~~Q #%99') A/* $D *,\W]WlWl)mD &?I r?s@r3rArALsN7 }7 37 c7 `c7 mp7 r6U\\6ell6r4rAc.eZdZdedeffd ZdZxZS) TextNetStagerdepthcpt ||j|}|j|}t |}|j |}|j |dz}|g|g|dz zz}|g|z} g} t || ||D]} | jt|g| tj| |_ y)Nr) rr conv_layer_kernel_sizesconv_layer_strideslen hidden_sizeszipappendrAr' ModuleListstage) r1rr\rr num_layersstage_in_channel_sizestage_out_channel_sizerBrCre stage_configr2s r3r zTextNetStage.__init__s 44U; **51% & 3 3E :!'!4!4UQY!?,-1G0HJYZN0[[ ./*<  \;O EL LL,VClC D E]]5) r4c8|jD] }||} |Sr8)re)r1 hidden_stateblocks r3r9zTextNetStage.forwards%ZZ /E .L /r4)r:r;r<rrYr r9r>r?s@r3r[r[s*}*S*"r4r[c beZdZdeffd Z ddej deedeede fdZ xZ S) TextNetEncoderrct|g}t|j}t |D]}|j t ||tj||_ yr8) rr r`r^rangercr[r'rdstages)r1rrq num_stagesstage_ixr2s r3r zTextNetEncoder.__init__s\ 778 j) :H MM,vx8 9 :mmF+ r4rkoutput_hidden_states return_dictr6c|g}|jD]}||}|j||s |f}|r||fzS|St||S)N)last_hidden_stater5)rqrcr )r1rkrtrur5reoutputs r3r9zTextNetEncoder.forwardse & [[ /E .L   . /"_F0D6],, P& P- \ijjr4rG) r:r;r<rr r=rrboolr r9r>r?s@r3rnrnsS,},04&* kllk'tnkd^ k ( kr4rnc&eZdZUeed<dZdZdZy)TextNetPreTrainedModelrtextnet pixel_valuesct|tjtjfrm|jj j d|jj|j%|jj jyyt|tjrW|jj jd|j%|jj jyyy)Ng)meanstdg?) r%r'Linearr(weightdatanormal_rinitializer_rangerzero_r,fill_)r1modules r3 _init_weightsz$TextNetPreTrainedModel._init_weightss fryy"))4 5 MM   & &CT[[5R5R & S{{&   &&('  / MM   $ $S ){{&   &&('0r4N)r:r;r<r__annotations__base_model_prefixmain_input_namerr4r3r{r{s !$O)r4r{creZdZfdZe ddedeedeedee e e e fe e e ffdZ xZS) TextNetModelct||t||_t ||_t jd|_|jy)N)rr) rr rstemrnencoderr'AdaptiveAvgPool2dpooler post_initr1rr2s r3r zTextNetModel.__init__sD  $V, %f- **62  r4r}rtrur6c:||n|jj}||n|jj}|j|}|j |||}|d}|j |}|s||f}|r ||dfzS|St |||r |dSdS)Nrtrurr)rw pooler_outputr5)ruse_return_dictrtrrrr) r1r}rtrurkencoder_outputsrw pooled_outputrxs r3r9zTextNetModel.forwards&1%r?s@r3rrsgos " :B4. ^fgk^l uS$s)^$eCj2ZZ [  r4rz TextNet Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for ImageNet. ) custom_introc eZdZfdZe ddeejdeejdee dee de f dZ xZ S) TextNetForImageClassificationct|||j|_t||_t j d|_t j|_ |jdkDr-t j|jd|jnt j|_ t j|j|jg|_|j!y)N)rrr)rr num_labelsrr|r'ravg_poolFlattenflattenrrar/fcrd classifierrrs r3r z&TextNetForImageClassification.__init__s   ++#F+ ,,V4 zz| KQK\K\_`K`"))F//3V5F5FGfhfqfqfs-- (EF r4r}labelsrtrur6c.||n|jj}|j|||}|d}|jD] }||} |j |}d} ||jj |j dk(rd|j_nl|j dkDrL|jtjk(s|jtjk(rd|j_nd|j_|jj dk(rIt} |j dk(r&| |j|j} n| ||} n|jj dk(r=t} | |jd|j |jd} n,|jj dk(rt} | ||} |s|f|d dz} | | f| zS| St!| ||j" S) al 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). Examples: ```python >>> import torch >>> import requests >>> from transformers import TextNetForImageClassification, TextNetImageProcessor >>> from PIL import Image >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = TextNetImageProcessor.from_pretrained("czczup/textnet-base") >>> model = TextNetForImageClassification.from_pretrained("czczup/textnet-base") >>> inputs = processor(images=image, return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) >>> outputs.logits.shape torch.Size([1, 2]) ```Nrrr regressionsingle_label_classificationmulti_label_classificationrr)losslogitsr5)rrr|rr problem_typerdtyper=longrYr squeezerviewrrr5) r1r}rrtruoutputsrwlayerrrloss_fctrxs r3r9z%TextNetForImageClassification.forward'sB&1%r?s@r3rrs 59-1/3&* Bsu001Bs))*Bs'tn Bs d^ Bs . BsBsr4rzP TextNet backbone, to be used with frameworks like DETR and MaskFormer. c `eZdZfdZe ddedeedeedee e e ffdZ xZ S)TextNetBackbonect||t| |t||_|j |_|jyr8)rr _init_backbonerr|rarErrs r3r zTextNetBackbone.__init__ssC   v&#F+ "// r4r}rtrur6c||n|jj}||n|jj}|j|d|}|r |jn|d}d}t |j D]\}}||jvs|||fz }|s |f} |r|r |jn|d}| |fz } | St||r|jdSddS)a Examples: ```python >>> import torch >>> import requests >>> from PIL import Image >>> from transformers import AutoImageProcessor, AutoBackbone >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = AutoImageProcessor.from_pretrained("czczup/textnet-base") >>> model = AutoBackbone.from_pretrained("czczup/textnet-base") >>> inputs = processor(image, return_tensors="pt") >>> with torch.no_grad(): >>> outputs = model(**inputs) ```NTrrr) feature_mapsr5 attentions) rrrtr|r5 enumerate stage_names out_featuresr ) r1r}rtrurr5ridxrerxs r3r9zTextNetBackbone.forward}s.&1%r?s@r3rrmsW os/ "/ :B4./ ^fgk^l/ uU|^+ ,/ / r4r)rrr{r)*rXtypingrrrr=torch.nnr'rrrr transformersr transformers.activationsr transformers.modeling_outputsr r rr1transformers.models.textnet.configuration_textnetrtransformers.utilsr!transformers.utils.backbone_utilsrutilsr get_loggerr:loggerModulerrAr[rnr{rrr__all__rr4r3rs.'' AA(, L&;#   H %.ryy.DW6"))W6t2990kRYYk: )_ ) ) " )" " J Rs$:Rs  Rsj ; ,m;  ; | ir4