o ߥi@snddlZddlmZdgZdZdededefddZd d d d d ddddedededededede fddZ ddde fddZ ddde fddZ ddde fddZ ddde fddZdd d!dd"d#ed$ed%ede fd&d'ZGd(d)d)ejZd d!dd*ded%ede fd+d,Zd!dd-d%ede fd.d/Zddde fd0d1Zddde fd2d3Zddde fd4d5ZdS)6Ntorchz%https://dl.fbaipublicfiles.com/dinov2 arch_name patch_sizereturncCs"|dddd}d||S)N_dinov2_)replace)rrcompact_arch_namer _/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/cv/anydoor/dinov2/hubconf.py_make_dinov2_model_namesr vit_largeig?mlpT)rimg_sizer init_values ffn_layer block_chunks pretrainedrrrrrc KsPddlm}t||} t|||||d} | jdi||j|di| } | S)N)vision_transformer)rrrrrr ) dinov2.modelsrrdictupdate__dict__) rrrrrrrkwargsvitsr vit_kwargsmodelr r r _make_dinov2_models r!)rcKtdd|d|S)zP DINOv2 ViT-S/14 model (optionally) pretrained on the LVD-142M dataset. vit_smallrrNr r!rrr r r dinov2_vits142 r'cKr")zC DINOv2 ViT-B/14 model pretrained on the LVD-142M dataset. vit_baser$Nr r%r&r r r dinov2_vitb14:r(r*cKr")zP DINOv2 ViT-L/14 model (optionally) pretrained on the LVD-142M dataset. rr$Nr r%r&r r r dinov2_vitl14Br(r+cKtddd|d|S)zP DINOv2 ViT-g/14 model (optionally) pretrained on the LVD-142M dataset. vit_giant2 swiglufusedrrrNr r%r&r r r dinov2_vitg14Jsr0ir model_name embed_dimlayersrr2r3r4c Ks|dvs Jd|td||d}|r>|dkrt|nd}td|d|d|d }tjj|d d }|j|d d |S)N)rrUnsupported number of layers: rirr/_linearz _head.pthcpu) map_locationF)strict)nnLinearstr_DINOV2_BASE_URLrhubload_state_dict_from_urlload_state_dict) r2r3r4rr linear_head layers_strurl state_dictr r r _make_dinov2_linear_headUsrFcs<eZdZdddejdejdeffddZdd ZZS) _LinearClassifierWrapperr)r4backbonerBr4cs t||_||_||_dS)N)super__init__rHrBr4)selfrHrBr4 __class__r r rJls  z!_LinearClassifierWrapper.__init__c Cs|jdkr$|j|}|dd}|dd}t||dg}nG|jdkrc|jj|ddd}t|ddd|ddd|ddd|d dd|d dddg}nJd |j||S) Nrx_norm_clstokenrx_norm_patchtokensrT)nreturn_class_tokenFr5) r4rHforward_featuressqueezercatmeanget_intermediate_layersrB)rKx cls_token patch_tokens linear_inputr r r forwardvs   0& z _LinearClassifierWrapper.forward) __name__ __module__ __qualname__r;ModuleintrJr] __classcell__r r rLr rGjs rGrr4rc KsHtd||d|}|j}|j}t||}t||||d}t|||dS)Nr$r1)rHrBr4r )r!r3rrrFrG) rr4rrrHr3rr2rBr r r _make_dinov2_linear_classifiers" re)r4rcKstdd||d|S)z Linear classifier (1 or 4 layers) on top of a DINOv2 ViT-S/14 backbone (optionally) pretrained on the LVD-142M dataset and trained on ImageNet-1k. r#rdNr re)r4rrr r r dinov2_vits14_lcs rgcKr")z Linear classifier (1 or 4 layers) on top of a DINOv2 ViT-B/14 backbone (optionally) pretrained on the LVD-142M dataset and trained on ImageNet-1k. r)r$Nr rfr&r r r dinov2_vitb14_lc rhcKr")z Linear classifier (1 or 4 layers) on top of a DINOv2 ViT-L/14 backbone (optionally) pretrained on the LVD-142M dataset and trained on ImageNet-1k. rr$Nr rfr&r r r dinov2_vitl14_lcrirjcKr,)z Linear classifier (1 or 4 layers) on top of a DINOv2 ViT-g/14 backbone (optionally) pretrained on the LVD-142M dataset and trained on ImageNet-1k. r-r.r/Nr rfr&r r r dinov2_vitg14_lcsrk)rtorch.nnr; dependenciesr>r=rbrfloatboolr!r'r*r+r0rFrarGrergrhrjrkr r r r sv