Uh@*dZddlmZmZmZmZmZddlZddlm Z ddlm Z ddlm Z m Z m Z ddlmZddlmZddlmZmZmZmZdd lmZdd lmZdd lmZd d lmZej>e Z!Gdde jDZ#Gdde jDZ$Gdde jDZ%Gdde jDZ&eGddeZ'eGdde'Z(edGdde'Z)edGdd e'eZ*gd!Z+y)"zPyTorch TextNet model.)AnyListOptionalTupleUnionN)Tensor)BCEWithLogitsLossCrossEntropyLossMSELoss)PreTrainedModel)ACT2CLS)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$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/catia.catastroantioquia-mas.com/valormas/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-r0r2)r3r7s r5forwardzTextNetConvLayer.forwardFs- -0  6 }--r6) __name__ __module__ __qualname__rr"torchrr; __classcell__r4s@r5rr*s,B}B6.U\\.ell.r6rc 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)rDrErrrr ) num_featuresepsNN)r!r" num_channelsrErrr)ReLUr&r* main_convr.r/main_batch_norm vertical_convvertical_batch_normhorizontal_convhorizontal_batch_norm rbr_identity) r3rrDrErrrvertical_paddinghorizontal_paddingr4s r5r"zTextNetRepConvLayer.__init__Vs '(& NQ&1,{1~/Aa.GH#%779 #%#   "~~Q !#')(^Q/( "D (*~~Q #%99') A/* $D *,\W]WlWl)mD &?I __doc__rintr"r?rr;r@rAs@r5rCrCLsN7 }7 37 c7 `c7 mp7 r6U\\6ell6r6rCc.eZdZdedeffd ZdZxZS) TextNetStagerdepthcpt ||j|}|j|}t |}|j |}|j |dz}|g|g|dz zz}|g|z} g} t || ||D]} | jt|g| tj| |_ y)Nr) r!r"conv_layer_kernel_sizesconv_layer_strideslen hidden_sizeszipappendrCr) ModuleListstage) r3rr^rr num_layersstage_in_channel_sizestage_out_channel_sizerDrErg stage_configr4s r5r"zTextNetStage.__init__s 44U; **51% & 3 3E :!'!4!4UQY!?,-1G0HJYZN0[[ ./*<  \;O EL LL,VClC D E]]5) r6c8|jD] }||} |Sr:)rg)r3 hidden_stateblocks r5r;zTextNetStage.forwards%ZZ /E .L /r6)r<r=r>rr[r"r;r@rAs@r5r]r]s*}*S*"r6r]c beZdZdeffd Z ddej deedeede fdZ xZ S) TextNetEncoderrct|g}t|j}t |D]}|j t ||tj||_ yr:) r!r"rbr`rangerer]r)rfstages)r3rrs num_stagesstage_ixr4s r5r"zTextNetEncoder.__init__s\ 778 j) :H MM,vx8 9 :mmF+ r6rmoutput_hidden_states return_dictr8c|g}|jD]}||}|j||s |f}|r||fzS|St||S)N)last_hidden_stater7)rsrer)r3rmrvrwr7rgoutputs r5r;zTextNetEncoder.forwardse & [[ /E .L   . /"_F0D6],, P& P- \ijjr6rI) r<r=r>rr"r?rrboolrr;r@rAs@r5rprpsS,},04&* kllk'tnkd^ k ( kr6rpceZdZeZdZdZdZy)TextNetPreTrainedModeltextnet pixel_valuesct|tjtjfrm|jj j d|jj|j%|jj jyyt|tjrW|jj jd|j%|jj jyyy)Ng)meanstdg?) r'r)Linearr*weightdatanormal_rinitializer_ranger zero_r.fill_)r3modules r5 _init_weightsz$TextNetPreTrainedModel._init_weightss fryy"))4 5 MM   & &CT[[5R5R & S{{&   &&('  / MM   $ $S ){{&   &&('0r6N)r<r=r>r config_classbase_model_prefixmain_input_namerr6r5r}r}s L!$O)r6r}creZdZfdZe ddedeedeedee e e e fe e e ffdZ xZS) TextNetModelct||t||_t ||_t jd|_|jy)N)rr) r!r"rstemrpencoderr)AdaptiveAvgPool2dpooler post_initr3rr4s r5r"zTextNetModel.__init__sD  $V, %f- **62  r6rrvrwr8c:||n|jj}||n|jj}|j|}|j |||}|d}|j |}|s||f}|r ||dfzS|St |||r |dSdS)Nrvrwrr)ry pooler_outputr7)ruse_return_dictrvrrrr) r3rrvrwrmencoder_outputsry pooled_outputrzs r5r;zTextNetModel.forwards&1%r"rrrr{rrrrrr;r@rAs@r5rrsgos " :B4. ^fgk^l uS$s)^$eCj2ZZ [  r6rz 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)r!r" num_labelsrr~r)ravg_poolFlattenflattenrrcr1fcrf classifierrrs r5r"z&TextNetForImageClassification.__init__s   ++#F+ ,,V4 zz| KQK\K\_`K`"))F//3V5F5FGfhfqfqfs-- (EF r6rlabelsrvrwr8c.||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)losslogitsr7)rrr~rr problem_typerdtyper?longr[r squeezer viewr rr7) r3rrrvrwoutputsrylayerrrloss_fctrzs r5r;z%TextNetForImageClassification.forward'sB&1%r"rrr? FloatTensor LongTensorr{rr;r@rAs@r5rrs 59-1/3&* Bsu001Bs))*Bs'tn Bs d^ Bs . BsBsr6rzP TextNet backbone, to be used with frameworks like DETR and MaskFormer. c `eZdZfdZe ddedeedeedee e e ffdZ xZ S)TextNetBackbonect||t| |t||_|j |_|jyr:)r!r"_init_backbonerr~rcrGrrs r5r"zTextNetBackbone.__init__ssC   v&#F+ "// r6rrvrwr8c||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_mapsr7 attentions) rrrvr~r7 enumerate stage_names out_featuresr) r3rrvrwrr7ridxrgrzs r5r;zTextNetBackbone.forward}s.&1%r"rrrr{rrrr;r@rAs@r5rrmsW os/ "/ :B4./ ^fgk^l/ uU|^+ ,/ / r6r)rrr}r),rZtypingrrrrrr?torch.nnr)rr r r transformersr transformers.activationsr transformers.modeling_outputsrrrr1transformers.models.textnet.configuration_textnetrtransformers.utilsr!transformers.utils.backbone_utilsrutilsr get_loggerr<loggerModulerrCr]rpr}rrr__all__rr6r5rs.44 AA(, L&;#   H %.ryy.DW6"))W6t2990kRYYk: )_ ) ) " )" " J Rs$:Rs  Rsj ; ,m;  ; | ir6