o hK@sdZddlmZddlZddlZddlmZmZddlmZm Z m Z ddl m Z ddl mZmZmZmZdd lmZdd lmZmZmZmZmZdd lmZd d lmZeeZ dZ!dZ"gdZ#dZ$dZ%dgZ&Gdddej'Z(Gdddej'Z)Gdddej'Z*Gdddej'Z+Gdddej'Z,Gdddej'Z-Gdddej'Z.Gd d!d!eZ/d"Z0d#Z1ed$e0Gd%d&d&e/Z2ed'e0Gd(d)d)e/Z3ed*e0Gd+d,d,e/eZ4dS)-z PyTorch ResNet model.)OptionalN)Tensornn)BCEWithLogitsLossCrossEntropyLossMSELoss)ACT2FN)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)add_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings) BackboneMixin) ResNetConfigrzmicrosoft/resnet-50)rirz tiger catc sHeZdZ ddededededef fd d Zd ed efd dZZS)ResNetConvLayerrrrelu in_channels out_channels kernel_sizestride activationcsVttj|||||ddd|_t||_|dur$t||_ dSt|_ dS)NF)rrpaddingbias) super__init__rConv2d convolution BatchNorm2d normalizationr Identityr)selfrrrrr __class__`/var/www/html/ai/venv/lib/python3.10/site-packages/transformers/models/resnet/modeling_resnet.pyr#?s  $zResNetConvLayer.__init__inputreturncCs"||}||}||}|SN)r%r'rr)r. hidden_stater,r,r-forwardIs   zResNetConvLayer.forward)rrr) __name__ __module__ __qualname__intstrr#rr3 __classcell__r,r,r*r-r>s rcs8eZdZdZdeffdd ZdedefddZZS) ResNetEmbeddingszO ResNet Embeddings (stem) composed of a single aggressive convolution. configcsBtt|j|jdd|jd|_tjdddd|_ |j|_dS)Nrr)rrrrr)rrr ) r"r#r num_channelsembedding_size hidden_actembedderr MaxPool2dpoolerr)r;r*r,r-r#Us  zResNetEmbeddings.__init__ pixel_valuesr/cCs4|jd}||jkrtd||}||}|S)NrzeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)shaper< ValueErrorr?rA)r)rCr< embeddingr,r,r-r3]s    zResNetEmbeddings.forward) r4r5r6__doc__rr#rr3r9r,r,r*r-r:Psr:csBeZdZdZd dededeffdd Zded efd d ZZS) ResNetShortCutz ResNet shortcut, used to project the residual features to the correct size. If needed, it is also used to downsample the input using `stride=2`. rrrrcs0ttj||d|dd|_t||_dS)NrF)rrr!)r"r#rr$r%r&r')r)rrrr*r,r-r#ns zResNetShortCut.__init__r.r/cCs||}||}|Sr0)r%r'r1r,r,r-r3ss  zResNetShortCut.forward)r) r4r5r6rGr7r#rr3r9r,r,r*r-rHhsrHc s<eZdZdZd dedededeffdd Zd d ZZS) ResNetBasicLayerzO A classic ResNet's residual layer composed by two `3x3` convolutions. rrrrrrcsft||kp |dk}|rt|||dnt|_tt|||dt||dd|_t ||_ dS)Nrrr r"r#rHrr(shortcut Sequentialrlayerr r)r)rrrrshould_apply_shortcutr*r,r-r#~s   zResNetBasicLayer.__init__cC.|}||}||}||7}||}|Sr0rOrMrr)r2residualr,r,r-r3    zResNetBasicLayer.forward)rr) r4r5r6rGr7r8r#r3r9r,r,r*r-rIys  rIc sLeZdZdZ    ddededed ed ed ef fd d ZddZZ 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. rrFrrrr reductiondownsample_in_bottleneckc st||kp |dk}||}|rt|||dnt|_tt||d|r)|nddt|||s3|nddt||ddd|_t ||_ dS)NrrJ)rr)rrrL) r)rrrrrXrYrPreduces_channelsr*r,r-r#s zResNetBottleNeckLayer.__init__cCrQr0rRrSr,r,r-r3rUzResNetBottleNeckLayer.forward)rrrWF) r4r5r6rGr7r8boolr#r3r9r,r,r*r-rVs( rVc sNeZdZdZ  ddededededef fdd Zd ed efd d ZZ S) ResNetStagez4 A ResNet stage composed by stacked layers. rr;rrrdepthcstjdkr tntjdkr||jjd}n ||jd}tj|gfddt |dDR|_ dS)N bottleneck)rrrY)rrcsg|] }jdqS)rK)r>).0_r;rOrr,r- sz(ResNetStage.__init__..r) r"r# layer_typerVrIr>rYrrNrangelayers)r)r;rrrr] first_layerr*rar-r#s  zResNetStage.__init__r.r/cCs|}|jD]}||}q|Sr0)re)r)r.r2rOr,r,r-r3s  zResNetStage.forward)rr) r4r5r6rGrr7r#rr3r9r,r,r*r-r\s  r\c s@eZdZdeffdd Z d dededed efd d ZZ S) ResNetEncoderr;c sttg|_|jt||j|jd|j rdnd|j ddt |j|jdd}t ||j ddD]\\}}}|jt||||dq9dS)Nrrr)rr])r]) r"r#r ModuleListstagesappendr\r= hidden_sizesdownsample_in_first_stagedepthszip)r)r;in_out_channelsrrr]r*r,r-r#s    zResNetEncoder.__init__FTr2output_hidden_states return_dictr/cCsb|rdnd}|jD] }|r||f}||}q |r||f}|s+tdd||fDSt||dS)Nr,css|] }|dur|VqdSr0r,)r_vr,r,r- sz(ResNetEncoder.forward..)last_hidden_state hidden_states)ritupler )r)r2rprqru stage_moduler,r,r-r3s     zResNetEncoder.forward)FT) r4r5r6rr#rr[r r3r9r,r,r*r-rgsrgc@s$eZdZdZeZdZdZddZdS)ResNetPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. resnetrCcCs^t|tjrtjj|jddddSt|tjtjfr-tj|jdtj|j ddSdS)Nfan_outr)mode nonlinearityrr) isinstancerr$initkaiming_normal_weightr& GroupNorm constant_r!)r)moduler,r,r- _init_weightss z#ResNetPreTrainedModel._init_weightsN) r4r5r6rGr config_classbase_model_prefixmain_input_namerr,r,r,r-rx s  rxaH This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`ResNetConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. aF Args: pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See [`ConvNextImageProcessor.__call__`] for details. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. zOThe bare ResNet model outputting raw features without any specific head on top.c s\eZdZfddZeeeeee de d d de de e de e d efd d ZZS) ResNetModelcs>t|||_t||_t||_td|_ | dS)N)rr) r"r#r;r:r?rgencoderrAdaptiveAvgPool2drA post_initrBr*r,r-r#<s     zResNetModel.__init__vision) checkpoint output_typermodalityexpected_outputNrCrprqr/cCs||dur|n|jj}|dur|n|jj}||}|j|||d}|d}||}|s6||f|ddSt|||jdS)Nrprqrr)rt pooler_outputru)r;rpuse_return_dictr?rrAr ru)r)rCrprqembedding_outputencoder_outputsrt pooled_outputr,r,r-r3Es    zResNetModel.forwardNN)r4r5r6r#rRESNET_INPUTS_DOCSTRINGr_CHECKPOINT_FOR_DOCr _CONFIG_FOR_DOC_EXPECTED_OUTPUT_SHAPErrr[r3r9r,r,r*r-r7s(  rz ResNet Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for ImageNet. c speZdZfddZeeeeee e d    d de e j de e jde ede ed ef d d ZZS) ResNetForImageClassificationcs^t||j|_t||_tt|jdkr#t|j d|jnt |_ | dS)Nr) r"r# num_labelsrryrrNFlattenLinearrkr( classifierrrBr*r,r-r#qs  $ z%ResNetForImageClassification.__init__)rrrrNrClabelsrprqr/c Csb|dur|n|jj}|j|||d}|r|jn|d}||}d}|dur|jjdurP|jdkr6d|j_n|jdkrL|jtj ksG|jtj krLd|j_nd|j_|jjdkrnt } |jdkrh| | | }n+| ||}n%|jjdkrt } | |d|j|d}n|jjdkrt} | ||}|s|f|dd} |dur|f| S| St|||jd S) a0 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). Nrr regressionsingle_label_classificationmulti_label_classificationrr)losslogitsru)r;rryrr problem_typerdtypetorchlongr7rsqueezerviewrr ru) r)rCrrprqoutputsrrrloss_fctoutputr,r,r-r3}s6    "       z$ResNetForImageClassification.forward)NNNN)r4r5r6r#rrr_IMAGE_CLASS_CHECKPOINTr r_IMAGE_CLASS_EXPECTED_OUTPUTrr FloatTensor LongTensorr[r3r9r,r,r*r-ris0  rzO ResNet backbone, to be used with frameworks like DETR and MaskFormer. c sVeZdZfddZeeeeed d de de e de e defd d Z Z S) ResNetBackbonecsHt|t||jg|j|_t||_t||_ | dSr0) r"r#_init_backboner=rk num_featuresr:r?rgrrrBr*r,r-r#s     zResNetBackbone.__init__)rrNrCrprqr/c Cs|dur|n|jj}|dur|n|jj}||}|j|ddd}|j}d}t|jD]\}} | |jvr;|||f7}q+|sK|f} |rI| |jf7} | St ||rU|jddSdddS)a3 Returns: Examples: ```python >>> from transformers import AutoImageProcessor, AutoBackbone >>> import torch >>> from PIL import Image >>> import requests >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> processor = AutoImageProcessor.from_pretrained("microsoft/resnet-50") >>> model = AutoBackbone.from_pretrained( ... "microsoft/resnet-50", out_features=["stage1", "stage2", "stage3", "stage4"] ... ) >>> inputs = processor(image, return_tensors="pt") >>> outputs = model(**inputs) >>> feature_maps = outputs.feature_maps >>> list(feature_maps[-1].shape) [1, 2048, 7, 7] ```NTrr,) feature_mapsru attentions) r;rrpr?rru enumerate stage_names out_featuresr ) r)rCrprqrrruridxstagerr,r,r-r3s0   zResNetBackbone.forwardr)r4r5r6r#rrrr rrrr[r3r9r,r,r*r-rs  r)5rGtypingrrtorch.utils.checkpointrrtorch.nnrrr activationsr modeling_outputsr r r r modeling_utilsrutilsrrrrrutils.backbone_utilsrconfiguration_resnetr get_loggerr4loggerrrrrr$RESNET_PRETRAINED_MODEL_ARCHIVE_LISTModulerr:rHrIrVr\rgrxRESNET_START_DOCSTRINGrrrrr,r,r,r-sZ      *&) .F