^jBdZddlZddlZddlmZmZddlmZddlm Z ddl m Z m Z ddl mZmZmZmZdd lmZdd lmZmZdd lmZd d lmZej6eZGddej<ZGddej<Z Gddej<Z!Gddej<Z"Gddej<Z#Gddej<Z$Gddej<Z%eGddeZ&eGdde&Z'ed !Gd"d#e&Z(ed$!Gd%d&e e&Z)gd'Z*y)(zPyTorch ResNet model.N)Tensornn)initialization)ACT2FN) BackboneMixinfilter_output_hidden_states)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)can_return_tuple) ResNetConfigceZdZ d dededeeeefzdededeeeefzdedeffd Zd ejd ejfd Z xZ S)ResNetConvLayer in_channels out_channels kernel_sizestridebiasdilationgroups activationc t |tj|||||dz||||_tj ||_|t||_ ytj|_ y)N)rrrrpaddingrrr) super__init__rConv2d convolution BatchNorm2d normalizationrIdentityr) selfrrrrrrrr __class__s u/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/resnet/modeling_resnet.pyr"zResNetConvLayer.__init__(sn 99#%#1$   ^^L90:0F&,BKKM hidden_statesreturncl|j|}|j|}|j|}|SN)r$r&r)r(r,s r*forwardzResNetConvLayer.forwardAs6((7 **=9  6 r+)rrFrrrelu) __name__ __module__ __qualname__inttupleboolstrr"torchrr0 __classcell__r)s@r*rr's ./*+ ZZZ5c?* Z  Z  Zc3h'ZZZ2U\\ellr+rc8eZdZdZdeffd ZdedefdZxZS)ResNetEmbeddingszO ResNet Embeddings (stem) composed of a single aggressive convolution. configct|t|j|jdd|j |_tjddd|_ |j|_y)Nr)rrrrr)rrr ) r!r"r num_channelsembedding_size hidden_actembedderr MaxPool2dpoolerr(r>r)s r*r"zResNetEmbeddings.__init__Ms\ '   !6!6Aa\b\m\m  llqAF "//r+ pixel_valuesr-c|jd}||jk7r td|j|}|j |}|S)NrzeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)shaperA ValueErrorrDrF)r(rHrA embeddings r*r0zResNetEmbeddings.forwardUsT#))!, 4,, ,w MM,/ KK * r+) r2r3r4__doc__rr"rr0r:r;s@r*r=r=Hs'0|0Fvr+r=cBeZdZdZd dededeffd ZdedefdZxZS) ResNetShortCutz ResNet shortcut, used to project the residual features to the correct size. If needed, it is also used to downsample the input using `stride=2`. rrrct|tj||d|d|_tj ||_y)NrF)rrr)r!r"rr#r$r%r&)r(rrrr)s r*r"zResNetShortCut.__init__fs: 99[,AV\chi^^L9r+inputr-cJ|j|}|j|}|Sr/)r$r&)r(rQ hidden_states r*r0zResNetShortCut.forwardks(''. )),7 r+)r) r2r3r4rMr5r"rr0r:r;s@r*rOrO`s5 :C:s:C: Vr+rOc <eZdZdZddedededeffd ZdZxZS) ResNetBasicLayerzO A classic ResNet's residual layer composed by two `3x3` convolutions. rrrrct|||k7xs|dk7}|rt|||ntj|_tj t|||t||d|_t||_ y)Nrrr r!r"rOrr'shortcut Sequentialrlayerrr)r(rrrrshould_apply_shortcutr)s r*r"zResNetBasicLayer.__init__vs{  +| ; Jv{H]N; V Dcecncncp ]] Kf E L,4 H  !,r+cz|}|j|}|j|}||z }|j|}|Sr/r\rZrr(rSresiduals r*r0zResNetBasicLayer.forwardAzz,/ ==* |4 r+)rr1) r2r3r4rMr5r8r"r0r:r;s@r*rUrUqs/ -C -s -C -Y\ -r+rUc LeZdZdZ d dedededededef fd Zd ZxZ S) ResNetBottleNeckLayera A classic ResNet's bottleneck layer composed by three `3x3` convolutions. The first `1x1` convolution reduces the input by a factor of `reduction` in order to make the second `3x3` convolution faster. The last `1x1` convolution remaps the reduced features to `out_channels`. If `downsample_in_bottleneck` is true, downsample will be in the first layer instead of the second layer. rrrr reductiondownsample_in_bottleneckc Ft |||k7xs|dk7}||z}|rt|||ntj|_tj t||d|r|ndt|||s|ndt||dd|_t||_ y)NrrW)rr)rrrY) r(rrrrrerfr]reduces_channelsr)s r*r"zResNetBottleNeckLayer.__init__s  +| ; Jv{'94H]N; V Dcecncncp ]] -1OgVmn  ,.>Umvst u ,lVZ [   !,r+cz|}|j|}|j|}||z }|j|}|Sr/r_r`s r*r0zResNetBottleNeckLayer.forwardrbr+)rr1F) r2r3r4rMr5r8r7r"r0r:r;s@r*rdrdsZ ).--- -  -  -#'-0r+rdc NeZdZdZ d dededededef fd Zded efd ZxZ S) ResNetStagez4 A ResNet stage composed by stacked layers. r>rrrdepthc t ||jdk(rtnt}|jdk(r"|||||j |j }n|||||j }tj|gt|dz Dcgc]}||||j c}|_ ycc}w)N bottleneck)rrrf)rrrrX) r!r" layer_typerdrUrCrfrr[rangelayers) r(r>rrrrmr\ first_layer_r)s r*r"zResNetStage.__init__s )/):):l)J%P`    ,!,,)/)H)H K \&U[UfUfgKmm  dijorsjsdtu_`5|HYHYZu  usB; rQr-c<|}|jD] }||} |Sr/)rr)r(rQrSr\s r*r0zResNetStage.forwards* [[ /E .L /r+)rr) r2r3r4rMrr5r"rr0r:r;s@r*rlrlsX          4Vr+rlc @eZdZdeffd Z ddedededefdZxZ S) ResNetEncoderr>c t|tjg|_|jj t ||j|jd|jrdnd|jdt|j|jdd}t||jddD]0\\}}}|jj t ||||2y)Nrrr)rrm)rm) r!r"r ModuleListstagesappendrlrB hidden_sizesdownsample_in_first_stagedepthszip)r(r>in_out_channelsrrrmr)s r*r"zResNetEncoder.__init__s mmB'   %%##A&"<z(ResNetEncoder.forward..sSqQ]Ss)last_hidden_stater,)rzr6r )r(rSrrr, stage_modules r*r0zResNetEncoder.forwardsv3  KK 6L# - ? ' 5L  6 )\O;MS\=$ASS S-*'  r+)FT) r2r3r4rr"rr7r r0r:r;s@r*rwrws=\|\$]a " :> UY ' r+rwcfeZdZUeed<dZdZdZddgZe jfdZ xZ S)ResNetPreTrainedModelr>resnetrH)imagerrOct||t|tjr#t j |jddyt|tjrt j|jtjd|jxtjj j|j\}}|dkDrdtj|z nd}t j|j| |yyd|j j"vrt j$|jt j&|jt j&|j(t j$|j*t-|d d t j&|j.yyy) Nfan_outr1)mode nonlinearity)arr BatchNormnum_batches_tracked)r! _init_weights isinstancerr#initkaiming_normal_weightLinearkaiming_uniform_mathsqrtrr9_calculate_fan_in_and_fan_outuniform_r)r2ones_zeros_ running_mean running_vargetattrr)r(modulefan_inrtboundr)s r*rz#ResNetPreTrainedModel._init_weights sE f% fbii (  YV T  *  ! !&--499Q< @{{&!HHMMGG V 17!DIIf-- fkkE659' F,,55 5 JJv}} % KK $ KK++ , JJv)) *v4d;G F667H 6r+) r2r3r4r__annotations__base_model_prefixmain_input_nameinput_modalities_no_split_modulesr9no_gradrr:r;s@r*rrs@  $O!*,<=U]]_88r+rc ReZdZfdZe ddededzdedzdefdZxZ S) ResNetModelct||||_t||_t ||_tjd|_ |jy)N)rr) r!r"r>r=rDrwencoderrAdaptiveAvgPool2drF post_initrGs r*r"zResNetModel.__init__$sK   (0 $V, **62  r+NrHrrr-c (||n|jj}||n|jj}|j|}|j |||}|d}|j |}|s ||f|ddzSt |||jS)Nrrrr)r pooler_outputr,)r>rrrDrrFr r,) r(rHrrkwargsembedding_outputencoder_outputsr pooled_outputs r*r0zResNetModel.forward-s%9$D $++JjJj &1% 1` a classification loss is computed (Cross-Entropy). Nrrr)losslogitsr,)r>rrrr loss_functionr r,) r(rHrrrroutputsrrroutputs r*r0z$ResNetForImageClassification.forwardas&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("microsoft/resnet-50") >>> model = AutoBackbone.from_pretrained( ... "microsoft/resnet-50", out_features=["stage1", "stage2", "stage3", "stage4"] ... ) >>> inputs = processor(image, return_tensors="pt") >>> outputs = model(**inputs) >>> feature_maps = outputs.feature_maps >>> list(feature_maps[-1].shape) [1, 2048, 7, 7] ```NTrr) feature_mapsr, attentions) r>rrrDrr, enumerate stage_names out_featuresr ) r(rHrrrrrr,ridxstagers r*r0zResNetBackbone.forwardsH&1%rsR &!H .,-.   H %biiBryy0RYY"ryy4'BII'T#"))#L& BII& R8O88:( '( ( V ,s#8,s  ,s^ K ]$9K  K \ er+