h<dZddlmZmZmZddlZddlmZddlmZddl m Z ddl m Z ddl mZmZmZmZddlmZdd lmZdd lmZd d lmZej4eZGd dej:ZGddej:ZGddej:Z Gddej:Z!eGdde Z"eGdde"Z#edGdde"Z$edGdde"eZ%gd Z&y)!zPyTorch TextNet model.)AnyOptionalUnionN)Tensor)PreTrainedModel)ACT2CLS)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$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 k/var/www/html/aiagenthome/venv/lib/python3.12/site-packages/transformers/models/textnet/modeling_textnet.pyrzTextNetConvLayer.__init__*s !22(( #)#7#7 &1159   "a '););A)>!)C D((A-  II  $ $  $ $//%%   ..)A)A6CXCXY++-  # # /%d&>&>?ADO 0 hidden_statesreturnch|j|}|j|}|j|SN)r(r+r-)r.r2s r0forwardzTextNetConvLayer.forwardEs- -0  6 }--r1) __name__ __module__ __qualname__r rtorchrr6 __classcell__r/s@r0rr)s,B}B6.U\\.ell.r1rc peZdZdZdededededef fd Zdejd ejfd Z xZ S) TextNetRepConvLayera This layer supports re-parameterization by combining multiple convolutional branches (e.g., main convolution, vertical, horizontal, and identity branches) during training. At inference time, these branches can be collapsed into a single convolution for efficiency, as per the re-parameterization paradigm. The "Rep" in the name stands for "re-parameterization" (introduced by RepVGG). r in_channels out_channelsrrctt |||_||_||_||_|ddz dz|ddz dzf}t j|_t j|||||d|_ t j||j|_ |ddz dzdf}d|ddz dzf}|ddk7rLt j|||ddf||d|_t j||j|_nd\|_|_|ddk7rLt j||d|df||d|_t j||j|_nd\|_|_||k(r,|dk(r't j||j|_yd|_y)NrrrF)r?r@rrrr) num_featuresepsNN)rr num_channelsr@rrr$ReLUr!r% main_convr)r*main_batch_norm vertical_convvertical_batch_normhorizontal_convhorizontal_batch_norm rbr_identity) r.rr?r@rrrvertical_paddinghorizontal_paddingr/s r0rzTextNetRepConvLayer.__init__Us '(& NQ&1,{1~/Aa.GH#%779 #%#   "~~Q !#')(^Q/( "D (*~~Q #%99') A/* $D *,\W]WlWl)mD &?I r>KsN7 }7 37 c7 `c7 mp7 r6U\\6ell6r1r>c.eZdZdedeffd ZdZxZS) TextNetStagerdepthcpt ||j|}|j|}t |}|j |}|j |dz}|g|g|dz zz}|g|z} g} t || ||D]} | jt|g| tj| |_ y)Nr) rrconv_layer_kernel_sizesconv_layer_strideslen hidden_sizeszipappendr>r$ ModuleListstage) r.rrYrr num_layersstage_in_channel_sizestage_out_channel_sizer?r@rb stage_configr/s r0rzTextNetStage.__init__s 44U; **51% & 3 3E :!'!4!4UQY!?,-1G0HJYZN0[[ ./*<  \;OL LL,VClC DP]]5) r1c8|jD] }||} |Sr5)rb)r. hidden_stateblocks r0r6zTextNetStage.forwards ZZE .L r1)r7r8r9r rVrr6r;r<s@r0rXrXs*}*S*"r1rXc beZdZdeffd Z ddej deedeede fdZ xZ S) TextNetEncoderrct|g}t|j}t |D]}|j t ||tj||_ yr5) rrr]r[ranger`rXr$rastages)r.rrn num_stagesstage_ixr/s r0rzTextNetEncoder.__init__sW 778 j)H MM,vx8 9*mmF+ r1rhoutput_hidden_states return_dictr3c|g}|jD]}||}|j||s |f}|r||fzS|St||S)N)last_hidden_stater2)rnr`r )r.rhrqrrr2rboutputs r0r6zTextNetEncoder.forwards` & [[E .L   .!"_F0D6],, P& P- \ijjr1rD) r7r8r9r rr:rrboolr r6r;r<s@r0rkrksS,},04&* kllk'tnkd^ k ( kr1rkc&eZdZUeed<dZdZdZy)TextNetPreTrainedModelrtextnet pixel_valuesct|tjtjfrm|jj j d|jj|j%|jj jyyt|tjrW|jj jd|j%|jj jyyy)Ng)meanstdg?) r"r$Linearr%weightdatanormal_rinitializer_rangerzero_r)fill_)r.modules r0 _init_weightsz$TextNetPreTrainedModel._init_weightss fryy"))4 5 MM   & &CT[[5R5R & S{{&   &&('  / MM   $ $S ){{&   &&('0r1N)r7r8r9r __annotations__base_model_prefixmain_input_namerr1r0rxrxs !$O)r1rxcreZdZfdZe ddedeedeedee e e e fe e e ffdZ xZS) TextNetModelct||t||_t ||_t jd|_|jy)N)rr) rrrstemrkencoderr$AdaptiveAvgPool2dpooler post_initr.rr/s r0rzTextNetModel.__init__sD  $V, %f- **62  r1rzrqrrr3c:||n|jj}||n|jj}|j|}|j |||}|d}|j |}|s||f}|r ||dfzS|St |||r |dSdS)Nrqrrrr)rt pooler_outputr2)ruse_return_dictrqrrrr ) r.rzrqrrrhencoder_outputsrt pooled_outputrus r0r6zTextNetModel.forwards&1% 1` a classification loss is computed (Cross-Entropy). Examples: ```python >>> import torch >>> import requests >>> from transformers import TextNetForImageClassification, TextNetImageProcessor >>> from PIL import Image >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = TextNetImageProcessor.from_pretrained("czczup/textnet-base") >>> model = TextNetForImageClassification.from_pretrained("czczup/textnet-base") >>> inputs = processor(images=image, return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) >>> outputs.logits.shape torch.Size([1, 2]) ```Nrrr)losslogitsr2)rrryrr loss_functionr r2) r.rzrrqrroutputsrtlayerrrrus r0r6z%TextNetForImageClassification.forward&sB&1%>> import torch >>> import requests >>> from PIL import Image >>> from transformers import AutoImageProcessor, AutoBackbone >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = AutoImageProcessor.from_pretrained("czczup/textnet-base") >>> model = AutoBackbone.from_pretrained("czczup/textnet-base") >>> inputs = processor(image, return_tensors="pt") >>> with torch.no_grad(): >>> outputs = model(**inputs) ```NTrrr) feature_mapsr2 attentions) rrrqryr2 enumerate stage_names out_featuresr ) r.rzrqrrrr2ridxrbrus r0r6zTextNetBackbone.forwardls.&1%rXrkrxrrr__all__rr1r0rs('' (, L&;#   H %.ryy.DW6"))W6t2990kRYYk: )_ ) ) " )" " J @s$:@s  @sF = ,m=  = @ ir1