g+  UddlmZddlZddlZddlmZddlmZmZm Z m Z ddl Z ddl Z ddl mZddlmZmZmZmZej(eZe dZeGdd Zd ed d gd gdddddd iZded< d ddZy)) annotationsN) dataclass)DictListLiteralOptional)forward_features_eva_fixedmake_vit_b16_backboneresize_patch_embed resize_vit) dinov2l16_384ceZdZUdZded<ded<dZded<dZded<d Zd ed <dZded <dZ ded <d Z ded< d Z ded<y ) ViTConfigzConfiguration for ViT.intin_chans embed_dimimg_size patch_sizeNz Optional[str] timm_preset timm_img_sizetimm_patch_sizez List[int]encoder_feature_layer_idsencoder_feature_dims) __name__ __module__ __qualname____doc____annotations__rrrrrrrJ/home/cameronsmith/repos/ml-depth-pro/src/depth_pro/network/vit_factory.pyrrsa MNHcJ"&K%M3OS,0y/Q&*)*Ur#rr) )ir&r&rrvit_large_patch14_dinov2i) rrrrrrrrrzDict[ViTPreset, ViTConfig]VIT_CONFIG_DICTct|}|j|jf}|j|jf}d|vrAtj|j |}t jt||_ n"tj|j |d}t||j|j|j|}|j|jk7rt|j ||_|j|j"k7rt%|j ||_|et'j(|d }|j+|d \} } t-| d k7rt/d | t-| d k7rt/d| t0j3||j S)aCreate and load a VIT backbone module. Args: ---- preset: The VIT preset to load the pre-defined config. use_pretrained: Load pretrained weights if True, default is False. checkpoint_uri: Checkpoint to load the wights from. use_grad_checkpointing: Use grandient checkpointing. Returns: ------- A Torch ViT backbone module. eva02) pretrainedT)r1dynamic_img_size)rr vit_featuresuse_grad_checkpointing)new_patch_size)rcpu) map_locationF) state_dictstrictrz(Found unexpected keys when loading vit: z#Keys are missing when loading vit: )r.rrtimm create_modelrtypes MethodTyper forward_featuresr rrrrr modelrr torchloadload_state_dictlenKeyErrorLOGGERinfo) presetuse_pretrainedcheckpoint_urir4configrrr?r8 missing_keysunexpected_keyss r$ create_vitrMDs(V $F1H##V%6%67J&!!&"4"4P!&!1!12Le!T!!   >D  " #88"("B"B%%5  EF222(ZP  &... x@ !ZZUC (-(=(=!%)>) % o  1 $EoEVWX X |  !@OP P KK ;;r#)FNF) rG ViTPresetrHboolrIz str | Noner4rOreturnz nn.Module) __future__rloggingr< dataclassesrtypingrrrrr:r@torch.nnnnvitr r r r getLoggerrrErNrr.r!rMr"r#r$rYs # !00    8 $    VV V,Y"13.  /+ "!!%#( 8 888! 8  8r#