a h|} @sUddlZddlmZddlmZddlmZmZmZm Z ddl Z ddl m Z ddl mZmZddlmZmZddlmZd d lmZmZmZd d lmZd d lmZmZgd ZGdddeZGdddeZ Gddde j!Z"Gddde j!Z#Gddde j!Z$e%e%e%e%e%e ee&ee$d ddZ'deiZ(e)e*efe+d<ie(dddZ,Gd d!d!eZ-Gd"d#d#eZ.Gd$d%d%eZ/Gd&d'd'eZ0Gd(d)d)eZ1eed*e-j2fd+dd,d-e e-e&ee$d.d/d0Z3eed*e.j2fd+dd,d-e e.e&ee$d.d1d2Z4eed*e/j2fd+dd,d-e e/e&ee$d.d3d4Z5eed*e0j2fd+dd,d-e e0e&ee$d.d5d6Z6eed7d+dd,d-e e1e&ee$d.d8d9Z7d@e%e%dd?Z8dS)AN) OrderedDict)partial)AnyCallable 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__nnZ BatchNorm2dr$rModuleZReLUr%r-r-S/var/www/auris/lib/python3.9/site-packages/torchvision/models/vision_transformer.pyr s r cs:eZdZdZdZeeedfdd ZfddZZ S)MLPBlockzTransformer MLP block.r)in_dimmlp_dimdropoutcsdtj|||gtjd|d|D]:}t|tjr$tj|j |j dur$tjj |j ddq$dS)N)r%Zinplacer2ư>std) super__init__r+ZGELUmodules isinstanceLinearinitZxavier_uniform_weightbiasnormal_)selfr0r1r2m __class__r-r.r7-s    zMLPBlock.__init__c s|dd}|dus|dkrxtdD]R} dD]H} |d| dd| } |d| d| } | |vr,|| || <q,q$t|||||||dS)Nversionr)r<r=Zlinear_r.)getrangepopr6_load_from_state_dict) r?Z state_dictprefixZlocal_metadatastrictZ missing_keysZunexpected_keysZ error_msgsrCitypeZold_keyZnew_keyrAr-r.rI6s"  zMLPBlock._load_from_state_dict) r&r'r(__doc___versionr)floatr7rI __classcell__r-r-rAr.r/(s r/c sZeZdZdZeejddfeeeeee de jj fdfdd Z e j dd d ZZS) EncoderBlockzTransformer encoder block.r3Zeps.) num_heads hidden_dimr1r2attention_dropoutr$csVt||_|||_tj|||dd|_t||_|||_ t ||||_ dS)NT)r2Z batch_first) r6r7rTln_1r+ZMultiheadAttentionself_attentionDropoutr2ln_2r/mlp)r?rTrUr1r2rVr$rAr-r.r7Ys    zEncoderBlock.__init__inputcCsjt|dkd|j||}|j|||dd\}}||}||}||}||}||S)NrE2Expected (batch_size, seq_length, hidden_dim) got F)Z need_weights) torch_assertdimshaperWrXr2rZr[)r?r]x_yr-r-r.forwardns    zEncoderBlock.forwardr&r'r(rNrr+ LayerNormr)rPrr_r,r7TensorrfrQr-r-rAr.rRVs rRc s^eZdZdZeejddfeeeeeeee de jj fdfdd Z e j dd d ZZS) Encoderz?Transformer Model Encoder for sequence to sequence translation.r3rS.) seq_length num_layersrTrUr1r2rVr$c s~tttd||jdd|_t||_ t } t |D] } t ||||||| d| <qBt | |_|||_dS)Nrg{Gz?r4Zencoder_layer_)r6r7r+ Parameterr_emptyr> pos_embeddingrYr2rrGrR Sequentiallayersln) r?rkrlrTrUr1r2rVr$rqrLrAr-r.r7}s    zEncoder.__init__r\cCs<t|dkd|j||j}||||S)NrEr^)r_r`rarbrorrrqr2)r?r]r-r-r.rfs zEncoder.forwardrgr-r-rAr.rjzs rjcseZdZdZddddeejdddfeeeeeeeeee ee de jj fe e ed fd d Ze je jd d d Ze jdddZZS)rz;Vision Transformer as per https://arxiv.org/abs/2010.11929.iNr3rS.) image_size patch_sizerlrTrUr1r2rV num_classesrepresentation_sizer$conv_stem_configsc  stt|t||dkd||_||_||_||_||_ ||_ | |_ | |_ | |_ | durt} d}t| D]:\}}| d|t||j|j|j|j |jd|j}qx| dtj||dd| |_ntjd|||d |_||d }ttdd||_|d7}t|||||||| |_||_t}| durRt || |d <n,t || |d <t!|d <t | | |d <t||_"t#|jtjr|jj$|jjd|jjd}tj%j&|jj't()d|d|jj*durbtj%+|jj*nj|jj,durbt#|jj,tjrbtj%j-|jj,j'dt()d|jj,jd|jj,j*durbtj%+|jj,j*t.|j"d rt#|j"j/tj r|j"j/j0}tj%j&|j"j/j't()d|dtj%+|j"j/j*t#|j"j1tj rtj%+|j"j1j'tj%+|j"j1j*dS)Nrz&Input shape indivisible by patch size!rEZ conv_bn_relu_) in_channelsr!r"r#r$r% conv_lastr)ryr!r")ryr!r"r#rhead pre_logitsZactr4rsg@)meanr5)2r6r7r r_r`rtrurUr1rVr2rvrwr$r+rp enumerateZ add_moduler r!r"r#r%ZConv2d conv_projrmzeros class_tokenrjencoderrkrr:ZTanhheadsr9ryr;Z trunc_normal_r<mathsqrtr=Zzeros_rzr>hasattrr|Z in_featuresr{)r?rtrurlrTrUr1r2rVrvrwr$rxZseq_projZ prev_channelsrLZconv_stem_layer_configrkZ heads_layersZfan_inrAr-r.r7s            zVisionTransformer.__init__)rcreturnc Cs|j\}}}}|j}t||jkd|jd|dt||jkd|jd|d||}||}||}|||j||}|ddd}|S)NzWrong image height! Expected z but got !zWrong image width! Expected rrr) rbrur_r`rtrreshaperUpermute) r?rcnchwpZn_hZn_wr-r-r._process_input s"" z VisionTransformer._process_input)rccCs^||}|jd}|j|dd}tj||gdd}||}|dddf}||}|S)Nrrra)rrbrexpandr_catrr)r?rcrZbatch_class_tokenr-r-r.rf!s    zVisionTransformer.forward)r&r'r(rNrr+rhr)rPrrr_r,listr r7rirrfrQr-r-rAr.rs.   ir) rurlrTrUr1weightsprogresskwargsrc Ks|durTt|dt|jd|jdd|jddks>Jt|d|jdd|dd}tf||||||d|} |r| |j|d d | S) Nrv categoriesmin_sizerrrt)rtrurlrTrUr1T)rZ check_hash)rlenmetarHrZload_state_dictZget_state_dict) rurlrTrUr1rrrrtmodelr-r-r._vision_transformer4s$    rr _COMMON_METAz(https://github.com/facebookresearch/SWAGz:https://github.com/facebookresearch/SWAG/blob/main/LICENSE)recipelicensec @seZdZedeeddiedddddd d id d d ddZedeeddej die dddddd idddddZ edeeddej die ddddddd id d dd dZ eZ d!S)"rz9https://download.pytorch.org/models/vit_b_16-c867db91.pthr crop_sizei(rrzNhttps://github.com/pytorch/vision/tree/main/references/classification#vit_b_16 ImageNet-1KgS㥛DT@g1ZW@zacc@1zacc@5gMb1@g(\t@ These weights were trained from scratch by using a modified version of `DeIT `_'s training recipe.  num_paramsrr_metrics_ops _file_size_docsurlZ 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_V1r BICUBIC_COMMON_SWAG_METAIMAGENET1K_SWAG_E2E_V1IMAGENET1K_SWAG_LINEAR_V1DEFAULTr-r-r-r.r_s rc @sHeZdZedeeddiedddddd 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 rrrr-r-r-r.rs( 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@rrNrr-r-r-r.rs rc @sHeZdZedeeddiedddddd d id d d ddZeZdS)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@rrrNrr-r-r-r.rs( rc @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 r rrrrrr-r-r-r.r2s`r pretrained)rT)rr)rrrrc Ks(t|}tfddddd||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:   rurlrTrUr1rr)rverifyrrrrr-r-r.rks rc Ks(t|}tfddddd||d|S)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: rrrr)rrrrr-r-r.rs rc Ks(t|}tfddddd||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: rr)rrrrr-r-r.rs rc Ks(t|}tfddddd||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: rrrrrr)rrrrr-r-r.rs r)rNc Ks(t|}tfddddd||d|S)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: rriir)rrrrr-r-r.rs rbicubicFzOrderedDict[str, torch.Tensor])rtru model_stateinterpolation_mode reset_headsrcCsh|d}|j\}}}|dkr,td|j||dd} | |krd|d8}| d8} |ddddddf} |ddddddf} | ddd} tt|} | | |krtd| | d|| d|| | } ||} tjj | | |d d }|d|| }|ddd}t j | |gdd }||d<|rdt }| 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: rNrzPseq_length is not a perfect square! Instead got seq_length_1d * seq_length_1d = z and seq_length = T)sizemodeZ align_cornersrr)rb ValueErrorrr)rrrr+Z functionalZ interpolater_rritems startswith)rtrurrrrorrkrUZnew_seq_lengthZpos_embedding_tokenZpos_embedding_imgZ seq_length_1dZnew_seq_length_1dZnew_pos_embedding_imgZnew_pos_embeddingZmodel_state_copykvr-r-r.interpolate_embeddingssF     r)rF)9r collectionsr functoolsrtypingrrrrr_Ztorch.nnr+Zops.miscr r Ztransforms._presetsr r utilsr Z_apirrrZ_metar_utilsrr__all__r r/r,rRrjrr)boolrrdictstrr*rrrrrrrrrrrrrr-r-r-r.s      .$& !OP9$ $ $ $ $$