o V h}@sUddlZddlmZddlmZddlmZmZmZm Z m Z m Z ddl Z ddl mZddlmZmZddlmZmZddlmZd d lmZmZmZd d lmZd d lmZmZgd Z Gddde Z!GdddeZ"Gdddej#Z$Gdddej#Z%Gdddej#Z&de'de'de'de'de'de ede(ded e&fd!d"Z)d#eiZ*ee+efe,d$<ie*d%d&d'Z-Gd(d)d)eZ.Gd*d+d+eZ/Gd,d-d-eZ0Gd.d/d/eZ1Gd0d1d1eZ2eed2e.j3fd3dd4d5de e.de(ded e&fd6d7Z4eed2e/j3fd3dd4d5de e/de(ded e&fd8d9Z5eed2e0j3fd3dd4d5de e0de(ded e&fd:d;Z6eed2e1j3fd3dd4d5de e1de(ded e&fdd3dd4d5de e2de(ded e&fd?d@Z8 A BdJdCe'de'dDdEdFe+dGe(d dEf dHdIZ9dS)KN) OrderedDict)partial)AnyCallableDictList NamedTupleOptional)Conv2dNormActivationMLP)ImageClassificationInterpolationMode)_log_api_usage_once)register_modelWeights WeightsEnum)_IMAGENET_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface) VisionTransformerViT_B_16_WeightsViT_B_32_WeightsViT_L_16_WeightsViT_L_32_WeightsViT_H_14_Weightsvit_b_16vit_b_32vit_l_16vit_l_32vit_h_14c@sVeZdZUeed<eed<eed<ejZedej fed<ej Z edej fed<dS)ConvStemConfig out_channels kernel_sizestride. norm_layeractivation_layerN) __name__ __module__ __qualname__int__annotations__nn BatchNorm2dr&rModuleReLUr'r1r1Y/var/www/vscode/kcb/lib/python3.10/site-packages/torchvision/models/vision_transformer.pyr" s r"cs>eZdZdZdZdededeffdd Zfdd ZZ S) MLPBlockzTransformer MLP block.r in_dimmlp_dimdropoutcsdtj|||gtjd|d|D]}t|tjr/tj|j |j dur/tjj |j ddqdS)N)r'inplacer6ư>std) super__init__r-GELUmodules isinstanceLinearinitxavier_uniform_weightbiasnormal_)selfr4r5r6m __class__r1r2r<-s   zMLPBlock.__init__c s|dd}|dus|dkr`_'s training recipe.  num_paramsrr_metrics_ops _file_size_docsurl transformsrz>https://download.pytorch.org/models/vit_b_16_swag-9ac1b537.pthr resize_size interpolationi^-)rrg~jtSU@giX@gˡEK@g|?5^t@ These weights are learnt via transfer learning by end-to-end fine-tuning the original `SWAG `_ weights on ImageNet-1K data. rrrrrrzAhttps://download.pytorch.org/models/vit_b_16_lc_swag-4e70ced5.pth+https://github.com/pytorch/vision/pull/5793gbX9xT@gQ X@ These weights are composed of the original frozen `SWAG `_ trunk weights and a linear classifier learnt on top of them trained on ImageNet-1K data. rrrrrrrNr(r)r*rrr r IMAGENET1K_V1rBICUBIC_COMMON_SWAG_METAIMAGENET1K_SWAG_E2E_V1IMAGENET1K_SWAG_LINEAR_V1DEFAULTr1r1r1r2r_s rc @HeZdZedeeddiedddddd d id d d ddZeZdS)rz9https://download.pytorch.org/models/vit_b_32-d86f8d99.pthrri1BrzNhttps://github.com/pytorch/vision/tree/main/references/classification#vit_b_32rg|?5^R@gW@rgA`Т@gl u@rrrN r(r)r*rrr rrrr1r1r1r2r* rc @seZdZedeedddieddddd d d id d dddZedeeddej die dddddd idddddZ edeeddej die ddddddd id d d d!dZ eZ d"S)#rz9https://download.pytorch.org/models/vit_l_16-852ce7e3.pthr)rri#rzNhttps://github.com/pytorch/vision/tree/main/references/classification#vit_l_16rg|?5^S@gFԨW@rgףp= N@g;O$@a These weights were trained from scratch by using a modified version of TorchVision's `new training recipe `_. rrz>https://download.pytorch.org/models/vit_l_16_swag-4f3808c9.pthri0)rrgjtV@gT㥛ĠX@gƟv@gy&11@rrzAhttps://download.pytorch.org/models/vit_l_16_lc_swag-4d563306.pthrgMbXIU@g^I [X@rrNrr1r1r1r2rs rc @r)rz9https://download.pytorch.org/models/vit_l_32-c7638314.pthrri[ErzNhttps://github.com/pytorch/vision/tree/main/references/classification#vit_l_32rg|?5>S@gGzDW@rgK7.@gE@rrrNrr1r1r1r2rrrc @seZdZedeeddejdiedddddd id d d d dZ edeeddejdieddddddd idddddZ e Z dS)rz>https://download.pytorch.org/models/vit_h_14_swag-80465313.pthri%)rrrgS#V@g#~jX@rg~jŏ@gK7I@rrrzAhttps://download.pytorch.org/models/vit_h_14_lc_swag-c1eb923e.pthrri@%rgZd;OmU@gQnX@g= ףpd@gI k֢@rrN) r(r)r*rrr rrrrrrr1r1r1r2r2sbr pretrained)rT)rrc K(t|}tdddddd||d|S)a Constructs a vit_b_16 architecture from `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale `_. Args: weights (:class:`~torchvision.models.ViT_B_16_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ViT_B_16_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.vision_transformer.VisionTransformer`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ViT_B_16_Weights :members:   rr~rdrer5rrNr1)rverifyrrrrr1r1r2rk rc Kr)a Constructs a vit_b_32 architecture from `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale `_. Args: weights (:class:`~torchvision.models.ViT_B_32_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ViT_B_32_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.vision_transformer.VisionTransformer`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ViT_B_32_Weights :members: rrrrNr1)rrrrr1r1r2rrrc Ks(t|}tdddddd||d|S)a Constructs a vit_l_16 architecture from `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale `_. Args: weights (:class:`~torchvision.models.ViT_L_16_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ViT_L_16_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.vision_transformer.VisionTransformer`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ViT_L_16_Weights :members: rrNr1)rrrrr1r1r2rrrc K(t|}tdddddd||d|S) a Constructs a vit_l_32 architecture from `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale `_. Args: weights (:class:`~torchvision.models.ViT_L_32_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ViT_L_32_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.vision_transformer.VisionTransformer`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ViT_L_32_Weights :members: rrrrrrNr1)rrrrr1r1r2r rr )rNc Kr) a Constructs a vit_h_14 architecture from `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale `_. Args: weights (:class:`~torchvision.models.ViT_H_14_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ViT_H_14_Weights` below for more details and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.vision_transformer.VisionTransformer`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ViT_H_14_Weights :members: rriirNr1)rrrrr1r1r2r!rr!bicubicFr model_statezOrderedDict[str, torch.Tensor]interpolation_mode reset_headscCs`|d}|j\}}}|dkrtd|j||dd} | |kr|d8}| d8} |ddddddf} |ddddddf} | ddd} tt|} | | |krdtd| | d|| d|| | } ||} tjj | | |d d }|d|| }|ddd}t j | |gdd }||d<|rt }| D] \}}|d s|||<q|}|S) aThis function helps interpolate positional embeddings during checkpoint loading, especially when you want to apply a pre-trained model on images with different resolution. Args: image_size (int): Image size of the new model. patch_size (int): Patch size of the new model. model_state (OrderedDict[str, torch.Tensor]): State dict of the pre-trained model. interpolation_mode (str): The algorithm used for upsampling. Default: bicubic. reset_heads (bool): If true, not copying the state of heads. Default: False. Returns: OrderedDict[str, torch.Tensor]: A state dict which can be loaded into the new model. zencoder.pos_embeddingrz%Unexpected position embedding shape: r NrzPseq_length is not a perfect square! Instead got seq_length_1d * seq_length_1d = z and seq_length = T)sizemode align_cornersrr)rt ValueErrorrr+rrrr- functional interpolaterqrritems startswith)rrrrrrrr}renew_seq_lengthpos_embedding_tokenpos_embedding_img seq_length_1dnew_seq_length_1dnew_pos_embedding_imgnew_pos_embeddingmodel_state_copykvr1r1r2interpolate_embeddingssH   r)rF):r collectionsr functoolsrtypingrrrrrr rqtorch.nnr-ops.miscr r transforms._presetsr rutilsr_apirrr_metar_utilsrr__all__r"r3r/rar|rr+boolrrstrr,rrrrrrrrrrr r!rr1r1r1r2s      .$& !OP9* * * * *$