+ jl`Rt^RIt^RIt^RIt^RIt^RIHtHt^RIH t ^RI H t ^RI HtHt^RIHtHtHtHt^RIHt^RIHtHt^R IHt^R IHt]P<!]4t R/R R llt!!R R]PD4t#!RR]PH4t%!RR]PL4t'!RR]PP4t)!RR]PL4t*!RR]PL4t+R0Rlt,!RR]PL4t-!RR]PL4t.!RR]PL4t/!R R!]PL4t0!R"R#]PL4t1]!R$R%]44t2]!R&R']244t3]!R(R)7!R*R+]244t4]!R,R)7!R-R.]]244t5.R1Ot6R#)2z9PyTorch BiT model. Also supports backbone for ViT hybrid.N)Tensornn)initialization)ACT2FN) BackboneMixinfilter_output_hidden_states)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)can_return_tuple BitConfigcFV^8dQhR\\\3,/#return)tuplebool)formats"e/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/bit/modeling_bit.py __annotate__r)s&&ERWY]R]L^&cRpVf(V^, W1^, ,,^,pW3#\V\4'dVP4pVR8XdRV^8XdCW1^, ,^,^8Xd(V^, W1^, ,,^,pW3#^pRpW3#VR8Xd^pW3#V^, W1^, ,,^,pW3#)a< Utility function to get the tuple padding value given the kernel_size and padding. Args: padding (Union[`str`, `int`], *optional*): Padding value, can be either `"same"`, `"valid"`. If a different value is provided the default padding from PyTorch is used. kernel_size (`int`, *optional*, defaults to 7): Kernel size of the convolution layers. stride (`int`, *optional*, defaults to 1): Stride value of the convolution layers. dilation (`int`, *optional*, defaults to 1): Dilation value of the convolution layers. FsameTvalid) isinstancestrlower)padding kernel_sizestridedilationdynamics&&&& rget_padding_valuer')sGQJ(Ao">>1D'3--/ f {!O <AQF"QJ(Ao*FF1L      G   h/&BBqHG  rc@aa]tRt^RtoRtRV3RlltRtRtVtV;t #)WeightStandardizedConv2dzConv2d with Weight Standardization. Used for ViT Hybrid model. Paper: [Micro-Batch Training with Batch-Channel Normalization and Weight Standardization](https://huggingface.co/papers/1903.10520) c <\WSWFR7wrZ\S V` VVVVVVVVR7V 'd\W4V4VnMRVnWnR#))r$r%)r$r"r%groupsbiasN)r'super__init__ DynamicPad2dpadeps) self in_channel out_channelsr#r$r"r%r+r,r1 is_dynamic __class__s &&&&&&&&&& rr.!WeightStandardizedConv2d.__init__Ys]0Vg       #KBDHDHrc VPeVPV4p\PPVPP ^VP R4RRRRVPR7PVP4p\PPWVPVPVPVPVP4pV#)NT)trainingmomentumr1)r0r functional batch_normweightreshaper4r1 reshape_asconv2dr,r$r"r%r+)r2 hidden_stater?s&& rforward WeightStandardizedConv2d.forwardvs 88 88L1L)) KK  4#4#4b 94PT_bhlhphp* *T[[ ! }}++ $))T[[$,, W[WbWb r)r1r0)SAMErFrFFgư> __name__ __module__ __qualname____firstlineno____doc__r.rD__static_attributes____classdictcell__ __classcell__r6 __classdict__s@@rr)r)Rs :  rr)c@aa]tRt^toRtRV3RlltRtRtVtV;t #)BitGroupNormActivationzI A module that combines group normalization with an activation function. c<\SV`VPW#VR7V'd\VP,VnR#\ P!4VnR#))r1affineN)r-r. num_groupsr hidden_act activationrIdentity)r2config num_channelsr1rVapply_activationr6s&&&&&&rr.BitGroupNormActivation.__init__s? **L&Q $V%6%67DO kkmDOrc\PPWPVPVP VP 4pVPV4pV#N)rr= group_normrWr?r,r1rY)r2rCs&&rrDBitGroupNormActivation.forwardsF}}// oot{{\`\e\egkgogop |4 r)rY)gh㈵>TTrHrQs@@rrTrTs,rrTc@aa]tRt^toRtRV3RlltRtRtVtV;t #)r/z A module that wraps dynamic padding of any input, given the parameters of the convolutional layer and the input hidden states. c<\SV`4\V\4'dW3p\V\4'dW"3p\V\4'dW33pWnW nW0nW@nRpWPnR#)c\\P!W, 4^, V,V^, V,,^,V, ^4#rF)maxmathceil)xr#r$r%s&&&&rcompute_padding.DynamicPad2d.__init__..compute_paddings@ !*-1V;{QRZ>ZZ]^^abbdef frN) r-r.rintr#r$r%valuerk)r2r#r$r%rnrkr6s&&&&& rr.DynamicPad2d.__init__sp  k3 ' '&4K fc " "%F h $ $ +H&    g /rc VP4RRwr#VPW P^,VP^,VP^,4pVPW0P^,VP^,VP^,4pV^8gV^8dY\ P PVV^,WU^,, V^,WD^,, .VPR7pV#)rN)rn) sizerkr#r$r%rr=r0rn)r2input input_height input_widthpadding_height padding_widths&& rrDDynamicPad2d.forwards$)JJL$5! --l)r#r$r1r")r#r$rNrGr9)r"rn preactivationr\)rFrFrFrF)r-r.r)r\embedding_sizeglobal_padding convolutionr|embedding_dynamic_paddingpoolerupperrrZr0 ConstantPad2d layer_typerTnormr2r[r6s&&rr.BitEmbeddings.__init__s 3     ! !))  #qPVPpPpq   ,1F1F1L1L1NRX1X{{}DH'' CHDH    /.vDYDYZDI DI"//rc&<V^8dQhRS[RS[/#)r pixel_valuesrr)rrRs"rrrsFvrcVP^,pW P8wd \R4hVPV4pVP V4pVP V4pVP V4pV#)rFzeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)shaper\ ValueErrorrr0rr)r2rr\ embeddings&& rrDBitEmbeddings.forwardsp#))!, ,, ,w $$\2 HHY' IIi( KK * r)rrr\r0rrHrQs@@rrrs#006rrcjaa]tRtRtoRtR V3RlV3RllltV3RlRltV3RlRltR tVt V;t #) BitDropPathizStochastic depth (DropPath) per sample, for residual blocks. Identity when ``drop_prob`` is 0 or outside training. See `Deep Networks with Stochastic Depth `_. c$<V^8dQhRS[RR/#)r drop_probrN)float)rrRs"rrBitDropPath.__annotate__s##%#$#rc0<\SV`4WnR#r`)r-r.r)r2rr6s&&rr.BitDropPath.__init__s "rcN<V^8dQhRS[PRS[P/#)rrr)torchr)rrRs"rrr s#<>S   & $$Q')DM4F4F4J,KK 50C0CML`L`a  M$=>   +m;;rc <V^8dQhRS[/#r)r )rrRs"rrr)s%%C%rc RVP 2#)zp=r)r2s&r extra_reprBitDropPath.extra_repr)sDNN#$$rr)r9) rIrJrKrLrMr.rDrrNrOrPrQs@@rrrs- ##<<%%%rrcTp\V\W^, ,4V,V,4pVRV,8d W1, pV#)rg?)rgrm)rndivisor min_value new_values&& rmake_divr-sCIIs5Q;#677BWLMI3; rc@aa]tRtRtoRtRV3RlltRtRtVtV;t #)BitPreActivationBottleneckLayeri5zPre-activation (v2) bottleneck block. Follows the implementation of "Identity Mappings in Deep Residual Networks": https://github.com/KaimingHe/resnet-1k-layers/blob/master/resnet-pre-act.lua Except it puts the stride on 3x3 conv when available. c ><\S V`4T;'gTpT;'gTp\W4,4p V 'd\VVVVRR7VnMRVn\ W4Vn\W+^RVPR7Vn \ WR7Vn \W^WXRVPR7Vn \ W4Vn \W^RVPR7Vn V ^8d\V 4VnR#\P !4VnR#)Tr$preactNrr1r"r)r$r+r1r")r-r.rBitDownsampleConv downsamplerTnorm1r)rconv1norm2conv2norm3conv3rrrZ drop_path) r2r[ in_channelsr4 bottle_ratior$r%first_dilationr+drop_path_rateis_first_layer mid_channelsr6s &&&&&&&&&&& rr.(BitPreActivationBottleneckLayer.__init__=s '338#22{  ;< / DO#DO+F@ -kPT^d^s^st +FN - &T[a[p[p  ,FA -l!QU_e_t_tu 8F8J^4PRP[P[P]rc:VPV4pTpVPeVPV4pVPV4pVPVP V44pVP VP V44pVPV4pW,#r`)rrrrrrrr)r2rhidden_states_preactshortcuts&& rrD'BitPreActivationBottleneckLayer.forwardis#zz-8! ?? &';F"a~55rc&<V^8dQhRS[RS[/#)rrsrr)rrRs"rrBitStage.__annotate__sVrcVTp\VP4F wr4V!V4pK V#r`) enumerater)r2rsrC_layers&& rrDBitStage.forwards, !$++.HA .L/r)r)rN) rIrJrKrLrMr.rrDrNrOrPrQs@@rrrs$,&\6 rrcZaa]tRtRtoV3RlV3RlltRtRV3RlRlltRtVtV;t #) BitEncoderi#c <V^8dQhRS[/#rr)rrRs"rrBitEncoder.__annotate__$s$:$:y$:rc <\SV`4\P!.4VnVP p^p^p\ P!\P!^VP\VP444PVP4Uu.uFpVP4NK pp\\!VPVP"V44F`wpwrp VP%WsWV4wrp\'VVV V VVV R7p T pW<,pVPP)\+V4V 4Kb R#uupi))r$r%rrN)r-r.r ModuleListstagesrrrnplinspacersumdepthssplittolistrzip hidden_sizesrrrr )r2r[rcurrent_strider%rjlayer_dropouts stage_idx current_depthcurrent_hidden_sizerr4r$stager6s&& rr.BitEncoder.__init__$s5 mmB' ((\\"++a1F1FFMMHZ"[\bbcicpcpq q HHJq  OX  v22N CO JIJ M.2-N-N+>&. *L(!#+E$H  $N KK " "3y>5 9+O  sE c\W5P,4pV^8Xd^M^pW%P8d WG,p^pWgV3#ry)r width_factor output_stride)r2rrrr%r[r4r$s&&&&&& rr'BitEncoder._get_updated_hyperparametersJsG 36I6I IJ 1n! 11 1  HFX--rc2<V^8dQhRS[RS[RS[RS[/#)rrCoutput_hidden_states return_dictr)rrr )rrRs"rrr Rs-  " :> UY ' rcV'dRMRpVPFpV'd WA3,pV!V4pK V'd WA3,pV'g/\;QJd.RW34FNK 5#!RW344#\VVR7#)Nc30"TF qfKVxK R#5ir`r).0vs& r %BitEncoder.forward..asS$Aq$As )last_hidden_staterr)rrr )r2rCr&r'r stage_modules&&&& rrDBitEncoder.forwardRs3  KKL# - ? ' 5L ( )O;M5S\$AS5 S5S\$ASS S-*'  r)r)FT) rIrJrKrLr.rrDrNrOrPrQs@@rr r #s$$:$:L.   rr cxaa]tRtRtoRtRtRtR.t]P!4V3Rl4t V3Rlt Rt Vt V;t#) BitPreTrainedModeliibitrrc|<\SV`V4\V\P4'd&\ P !VPRRR7R#\V\P4'd\ P!VP\P!^4R7VPe\PP PVP4wr#V^8d^\P!V4, M^p\ P!VPV)V4R#R#R#)fan_outrelu)mode nonlinearity)aN)r- _init_weightsrrConv2dinitkaiming_normal_r?Linearkaiming_uniform_rhsqrtr,r_calculate_fan_in_and_fan_outuniform_)r2modulefan_inrboundr6s&& rr: BitPreTrainedModel._init_weightsqs f% fbii ( (  YV T  * *  ! !&--499Q< @{{&!HHMMGG V 17!DIIf-- fkkE659'+rc&<V^8dQh/S[;R&#rr)rrRs"rrBitPreTrainedModel.__annotate__is rr)image)rIrJrKrLbase_model_prefixinput_modalitiesmain_input_name_no_split_modulesrno_gradr:__annotate_func__rNrOrPrQs@@rr2r2is@!$O() ]]_ : :rr2cRaa]tRtRtoV3Rlt]RV3RlRll4tRtVtV;t #)BitModelicX<\SV`V4Wn\V4Vn\ V4VnVPR8Xd\WPR,R7M\P!4Vn \P!R4VnVP4R#)rrNr<)rFrF)r-r.r[rembedderr encoderrrTrrrZrAdaptiveAvgPool2dr post_initrs&&rr.BitModel.__init__s   %f- !&)   O3 #68K8KB8O P **62  rcN<V^8dQhRS[RS[R,RS[R,RS[/#rrr&Nr'r)rrr )rrRs"rrBitModel.__annotate__s;   #Tk D[  2  rc rVeTMVPPpVeTMVPPpVPV4pVP WRVR7pV^,pVP V4pVP V4pV'gWx3VR,,#\VVVPR7#)Nr&r':rFNN)r. pooler_outputr) r[r&r'rSrTrrr r) r2rr&r'kwargsembedding_outputencoder_outputsr. pooled_outputs &&&&, rrDBitModel.forwards%9$D $++JjJj &1%rrZ classifierrVrs&&rr."BitForImageClassification.__init__s   ++F#-- JJLEKEVEVYZEZBIIf))"-v/@/@ A`b`k`k`m  rc <V^8dQhRS[PR,RS[PR,RS[R,RS[R,RS[/#)rrNlabelsr&r'r)r FloatTensor LongTensorrr )rrRs"rr&BitForImageClassification.__annotate__sbss''$.s  4's#Tk s D[ s .src VeTMVPPpVPWVR7pV'd VPMV^,pVP V4pRp VeVP W(VP4p V'g"V3VR,,p V e V 3V ,#T #\ WVPR7#)a 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 classification loss is computed (Cross-Entropy). Nr\:rNN)losslogitsr)r[r'r3r]rk loss_functionr r) r2rrnr&r'r^outputsrartrsoutputs &&&&&, rrD!BitForImageClassification.forwards&1%>> from transformers import AutoImageProcessor, AutoBackbone >>> import torch >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> processor = AutoImageProcessor.from_pretrained("google/bit-50") >>> model = AutoBackbone.from_pretrained("google/bit-50") >>> inputs = processor(image, return_tensors="pt") >>> outputs = model(**inputs) ```NTr\) feature_mapsr attentionsr) r[r'r&r3rr stage_names out_featuresr) r2rr&r'r^rvrridxrrws &&&&, rrDBitBackbone.forwards<&1%rs@  &!H .,-(   H %&R-ryy-`R\\$02990f 2<< 6/BII/f%"))%0A(biiA(HFFR' '.GryyGTC C L:::*2 !2 2 j ,s 2,s  ,s^ B -!3B  B J Yr