`i<8dZddlmZmZmZddlZddlmZddlmZddl m Z ddl m Z ddl mZmZmZmZddlmZdd lmZdd lmZd d lmZejeZGd dejZGddejZGddejZ GddejZ!eGdde Z"eGdde"Z#edGdde"Z$edGdde"eZ%gd Z&dS)!zPyTorch TextNet model.)AnyOptionalUnionN)Tensor)PreTrainedModel)ACT2CLS)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention) TextNetConfig)logging) BackboneMixin)auto_docstringcHeZdZdeffd ZdejdejfdZxZS)TextNetConvLayerconfigcDt|j|_|j|_|j|_t|jtr |jddz|jddzfn |jdz}tj |j |j |j|j|d|_tj|j |j|_tj|_|j t)|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__s /home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/transformers/models/textnet/modeling_textnet.pyrzTextNetConvLayer.__init__*s !2( #)#7 &1599 .V  "a ');A)>!)C D D(A-  I  $  $/%     .)A6CXYY+--  # /%d&>?AADOOO 0 / hidden_statesreturnc||}||}||SN)r(r+r-)r.r2s r0forwardzTextNetConvLayer.forwardEs6 -00  66 }---r1) __name__ __module__ __qualname__r rtorchrr6 __classcell__r/s@r0rr)sqB}BBBBBB6.U\.el........r1rc \eZdZdZdededededef fd Zdejd ejfd Z xZ S) TextNetRepConvLayera This layer supports re-parameterization by combining multiple convolutional branches (e.g., main convolution, vertical, horizontal, and identity branches) during training. At inference time, these branches can be collapsed into a single convolution for efficiency, as per the re-parameterization paradigm. The "Rep" in the name stands for "re-parameterization" (introduced by RepVGG). r in_channels out_channelsrrct||_||_||_||_|ddz dz|ddz dzf}t j|_t j |||||d|_ t j ||j |_ |ddz dzdf}d|ddz dzf}|ddkrHt j |||ddf||d|_t j ||j |_nd\|_|_|ddkrHt j ||d|df||d|_t j ||j |_nd\|_|_||kr!|dkrt j ||j nd|_dS)NrrrF)r?r@rrrr) num_featuresepsNN)rr num_channelsr@rrr$ReLUr!r% main_convr)r*main_batch_norm vertical_convvertical_batch_normhorizontal_convhorizontal_batch_norm rbr_identity) r.rr?r@rrrvertical_paddinghorizontal_paddingr/s r0rzTextNetRepConvLayer.__init__Us  '(& NQ&1,{1~/Aa.GH#%799 #%#     "~Q  !#')(^Q/( """D (*~Q  #%9') A/* $$$D *,\W]Wl)m)m)mD & &?I r>Ks7 }7 37 c7 `c7 mp7 7 7 7 7 7 r6U\6el66666666r1r>c.eZdZdedeffd ZdZxZS) TextNetStagerdepthct|j|}|j|}t |}|j|}|j|dz}|g|g|dz zz}|g|z} g} t || ||D]"} | t|g| R#tj | |_ dS)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 4U; *51%% & 3E :!'!4UQY!?,-1G0HJYZN0[[ ./*<  \;OO E EL LL,VClCCC D D D D]5)) r1c0|jD] }||}|Sr5)rb)r. hidden_stateblocks r0r6zTextNetStage.forwards*Z / /E 5..LLr1)r7r8r9r rVrr6r;r<s@r0rXrXsZ*}*S******"r1rXc deZdZdeffd Z d dejdeedeede fdZ xZ S) TextNetEncoderrctg}t|j}t |D]%}|t ||&tj||_ dSr5) rrr]r[ranger`rXr$rastages)r.rrn num_stagesstage_ixr/s r0rzTextNetEncoder.__init__sy 788 j)) : :H MM,vx88 9 9 9 9mF++ r1Nrhoutput_hidden_states return_dictr3c|g}|jD]"}||}||#|s |f}|r||fzn|St||S)N)last_hidden_stater2)rnr`r )r.rhrqrrr2rboutputs r0r6zTextNetEncoder.forwards~ & [ / /E 5..L   . . . . Q"_F0DP6],,,& P- \ijjjjr1rD) r7r8r9r rr:rrboolr r6r;r<s@r0rkrks,},,,,,,04&* kklk'tnkd^ k ( kkkkkkkkr1rkc(eZdZUeed<dZdZdZdS)TextNetPreTrainedModelrtextnet pixel_valuesct|tjtjfrT|jjd|jj|j |j j dSdSt|tj rF|jj d|j "|j j dSdSdS)Ng)meanstdg?) r"r$Linearr%weightdatanormal_rinitializer_rangerzero_r)fill_)r.modules r0 _init_weightsz$TextNetPreTrainedModel._init_weightss fry")4 5 5 ) M  & &CT[5R & S S S{&  &&((((('&  / / ) M  $ $S ) ) ){&  &&((((( ) )&&r1N)r7r8r9r __annotations__base_model_prefixmain_input_namerr1r0rxrxs< !$O)))))r1rxceZdZfdZe ddedeedeedee e e e fe e e ffdZ xZS) TextNetModelct|t||_t ||_t jd|_| dS)N)rr) rrrstemrkencoderr$AdaptiveAvgPool2dpooler post_initr.rr/s r0rzTextNetModel.__init__sa    $V,, %f-- *622  r1Nrzrqrrr3c:||n |jj}||n |jj}||}||||}|d}||}|s||f}|r ||dfzn|St |||r|dndS)Nrqrrrr)rt pooler_outputr2)ruse_return_dictrqrrrr ) r.rzrqrrrhencoder_outputsrt pooled_outputrus r0r6zTextNetModel.forwards&1% 1` a classification loss is computed (Cross-Entropy). Examples: ```python >>> import torch >>> import requests >>> from transformers import TextNetForImageClassification, TextNetImageProcessor >>> from PIL import Image >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = TextNetImageProcessor.from_pretrained("czczup/textnet-base") >>> model = TextNetForImageClassification.from_pretrained("czczup/textnet-base") >>> inputs = processor(images=image, return_tensors="pt") >>> with torch.no_grad(): ... outputs = model(**inputs) >>> outputs.logits.shape torch.Size([1, 2]) ```Nrrr)losslogitsr2)rrryrr loss_functionr r2) r.rzrrqrroutputsrtlayerrrrus r0r6z%TextNetForImageClassification.forward&sB&1%>> import torch >>> import requests >>> from PIL import Image >>> from transformers import AutoImageProcessor, AutoBackbone >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = AutoImageProcessor.from_pretrained("czczup/textnet-base") >>> model = AutoBackbone.from_pretrained("czczup/textnet-base") >>> inputs = processor(image, return_tensors="pt") >>> with torch.no_grad(): >>> outputs = model(**inputs) ```NTrrr) feature_mapsr2 attentions) rrrqryr2 enumerate stage_names out_featuresr ) r.rzrqrrrr2ridxrbrus r0r6zTextNetBackbone.forwardls.&1%rXrkrxrrr__all__rr1r0rs '''''''''' ((((((,,,,,, LKKKKK&&&&&&;;;;;;######  H % %.....ry...DW6W6W6W6W6")W6W6W6t290kkkkkRYkkk: ) ) ) ) )_ ) ) ) " " " " " )" " " J @s@s@s@s@s$:@s@s  @sF = = = = = ,m= =  = @ i h hr1