o wi@@shdZddlmZmZmZddlZddlmZddlmZddlm Z m Z m Z ddl m Z ddlmZddlmZmZmZmZdd lmZdd lmZdd lmZd d lmZeeZGdddej Z!Gdddej Z"Gdddej Z#Gdddej Z$eGddde Z%eGddde%Z&eddGddde%Z'eddGdd d e%eZ(gd!Z)dS)"zPyTorch TextNet model.)AnyOptionalUnionN)Tensor)BCEWithLogitsLossCrossEntropyLossMSELoss)PreTrainedModel)ACT2CLS)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention) TextNetConfig)logging) BackboneMixin)auto_docstringcs8eZdZdeffdd ZdejdejfddZZS)TextNetConvLayerconfigcst|j|_|j|_|j|_t|jt r%|jdd|jddfn|jd}t j |j |j |j|j|dd|_t |j |j|_t |_|jdurVt|j|_dSdS)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__i/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/textnet/modeling_textnet.pyr+s*    zTextNetConvLayer.__init__ hidden_statesreturncCs||}||}||SN)r(r+r-)r.r3r1r1r2forwardFs   zTextNetConvLayer.forward) __name__ __module__ __qualname__rrtorchrr6 __classcell__r1r1r/r2r*src sLeZdZdZdededededef fdd Zd ejd ejfd d Z Z 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_channelsrrc sft||_||_||_||_|ddd|dddf}t|_tj |||||dd|_ tj ||j d|_ |ddddf}d|dddf}|ddkrotj |||ddf||dd|_tj ||j d|_nd\|_|_|ddkrtj ||d|df||dd|_tj ||j d|_nd\|_|_||kr|dkrtj ||j d|_dSd|_dS)NrrrF)r=r>rrrr) num_featuresepsNN)rr num_channelsr>rrr$ReLUr!r% main_convr)r*main_batch_norm vertical_convvertical_batch_normhorizontal_convhorizontal_batch_norm rbr_identity) r.rr=r>rrrvertical_paddinghorizontal_paddingr/r1r2rVsZ         zTextNetRepConvLayer.__init__r3r4cCs||}||}|jdur||}||}||}|jdur0||}||}||}|jdur>||}||}||Sr5)rDrErFrGrHrIrJr!)r.r3 main_outputsvertical_outputshorizontal_outputsid_outr1r1r2r6s           zTextNetRepConvLayer.forward) r7r8r9__doc__rintrr:rr6r;r1r1r/r2r<Ls" 9r<cs.eZdZdedeffdd ZddZZS) TextNetStagerdepthc st|j|}|j|}t|}|j|}|j|d}|g|g|d}|g|} g} t|| ||D] } | t|g| Rq7t | |_ dS)Nr) rrconv_layer_kernel_sizesconv_layer_strideslen hidden_sizeszipappendr<r$ ModuleListstage) r.rrTrr num_layersstage_in_channel_sizestage_out_channel_sizer=r>r\ stage_configr/r1r2rs     zTextNetStage.__init__cCs|jD]}||}q|Sr5)r\)r. hidden_stateblockr1r1r2r6s  zTextNetStage.forward)r7r8r9rrRrr6r;r1r1r/r2rSsrSc sLeZdZdeffdd Z  d dejdeedeede fd d Z Z S) TextNetEncoderrcsFtg}t|j}t|D] }|t||qt||_ dSr5) rrrWrUrangerZrSr$r[stages)r.rre num_stagesstage_ixr/r1r2rs   zTextNetEncoder.__init__Nraoutput_hidden_states return_dictr4cCsL|g}|jD] }||}||q|s |f}|r||fS|St||dS)N)last_hidden_stater3)rerZr )r.rarhrir3r\outputr1r1r2r6s   zTextNetEncoder.forwardrA) r7r8r9rrr:rrboolr r6r;r1r1r/r2rcs rcc@s eZdZeZdZdZddZdS)TextNetPreTrainedModeltextnet pixel_valuescCst|tjtjfr#|jjjd|jjd|j dur!|j j dSdSt|tj r=|jj d|j dur?|j j dSdSdS)Ng)meanstdg?) r"r$Linearr%weightdatanormal_rinitializer_rangerzero_r)fill_)r.moduler1r1r2 _init_weightss   z$TextNetPreTrainedModel._init_weightsN)r7r8r9r config_classbase_model_prefixmain_input_namerzr1r1r1r2rms  rmcs`eZdZfddZe d dedeedeedee e e e fe e e ffdd Z ZS) TextNetModelcs8t|t||_t||_td|_| dS)N)rr) rrrstemrcencoderr$AdaptiveAvgPool2dpooler post_initr.rr/r1r2rs     zTextNetModel.__init__Nrorhrir4c Cs|dur|n|jj}|dur|n|jj}||}|j|||d}|d}||}|s;||f}|r9||dfS|St|||rF|ddSddS)Nrhrirr)rj pooler_outputr3)ruse_return_dictrhrrrr ) r.rorhriraencoder_outputsrj pooled_outputrkr1r1r2r6s&   zTextNetModel.forwardrA)r7r8r9rrrrrlrr#rlistr r6r;r1r1r/r2r~s r~z TextNet Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for ImageNet. ) custom_introc s\eZdZfddZe    d deejdeejdee dee de f d d Z Z S) TextNetForImageClassificationcs|t||j|_t||_td|_t|_ |jdkr)t |j d|jnt |_ t|j|j g|_|dS)N)rrr)rr num_labelsr~rnr$ravg_poolFlattenflattenrrrXr,fcr[ classifierrrr/r1r2rs    ( z&TextNetForImageClassification.__init__Nrolabelsrhrir4c Csl|dur|n|jj}|j|||d}|d}|jD]}||}q||}d} |dur|jjdurU|jdkr;d|j_n|jdkrQ|jtj ksL|jtj krQd|j_nd|j_|jjdkrst } |jdkrm| | | } n+| ||} n%|jjdkrt } | |d|j|d} n|jjdkrt} | ||} |s|f|d d} | dur| f| S| St| ||jd S) al labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the image classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Examples: ```python >>> import torch >>> import requests >>> from transformers import TextNetForImageClassification, TextNetImageProcessor >>> from PIL import Image >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = TextNetImageProcessor.from_pretrained("czczup/textnet-base") >>> model = TextNetForImageClassification.from_pretrained("czczup/textnet-base") >>> inputs = processor(images=image, return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) >>> outputs.logits.shape torch.Size([1, 2]) ```Nrrr regressionsingle_label_classificationmulti_label_classificationrr)losslogitsr3)rrrnrr problem_typerdtyper:longrRrsqueezerviewrrr3) r.rorrhrioutputsrjlayerrrloss_fctrkr1r1r2r6's:!      "       z%TextNetForImageClassification.forward)NNNN) r7r8r9rrrr: FloatTensor LongTensorrlrr6r;r1r1r/r2rs$ rzP TextNet backbone, to be used with frameworks like DETR and MaskFormer. c sReZdZfddZe d dedeedeedee e e ffdd Z Z S) TextNetBackbonecs6t|t|t||_|j|_|dSr5)rr_init_backboner~rnrXr?rrr/r1r2rss    zTextNetBackbone.__init__Nrorhrir4c Cs|dur|n|jj}|dur|n|jj}|j|d|d}|r!|jn|d}d}t|jD]\}}||jvr<|||f7}q,|sT|f} |rR|rI|jn|d}| |f7} | St||r^|jddSdddS)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) ```NTrrr1) feature_mapsr3 attentions) rrrhrnr3 enumerate stage_names out_featuresr ) r.rorhrirr3ridxr\rkr1r1r2r6}s0  zTextNetBackbone.forwardrA) r7r8r9rrrrrlrr#r r6r;r1r1r/r2rms  r)rr~rmr)*rQtypingrrrr:torch.nnr$rrrr transformersr transformers.activationsr transformers.modeling_outputsr r r r1transformers.models.textnet.configuration_textnetrtransformers.utilsr!transformers.utils.backbone_utilsrutilsr get_loggerr7loggerModulerr<rSrcrmr~rr__all__r1r1r1r2s>         "Z%U >