ehJddlZddlmZddZGddejZy)NcJtjj|||ddS)NbilinearF)size scale_factormode align_corners)nn functional interpolate)imagerrs C/home/cameronsmith/repos/canon_controll3r/old/feature_extractors.pyresizers- == $ $ !  % c0eZdZ dfd ZddZxZS) SpatialDinoct|tjj d||_|j j |_||_||_ |r'|j jD] }d|_ yy)Nzfacebookresearch/dinov2F) super__init__torchhubloadmodel embed_dim feature_dim num_patches_x num_patches_y parameters requires_grad)selffreeze_weights model_typerrparam __class__s r rzSpatialDino.__init__su YY^^$=zJ :://** ..0 ,&+# , rc|j^}}}}|jd|||}|jj|d}|j ddd}|jd|j |dz|dz}|r#t ||j|jf}|jg||j |j|j}|S)a Spatial dimensions of output will be H // 14, W // 14. If autoresize is True, then the output will be resized to the correct dimensions. Args: x (torch.Tensor): Images (B, C, H, W). Should be ImageNet normalized. autoresize (bool): Whether to resize the input to match the num_patch dimensions. Returns: feature_map (torch.tensor): (B, C, h, w) x_norm_patchtokensr)r) shapereshaperforward_featurespermuterrrr)rx autoresizeBchwfeaturess r forwardzSpatialDino.forward"sgg Aq! IIb!Q "::..q12FG##Aq!,##b$*:*:AGQ"WN hd.@.@$BTBT-UVH#8##     "&"4"4 6:6H6H r)T dinov2_vits14r7)F)__name__ __module__ __qualname__rr5 __classcell__)r#s@r rrs" ,"rr)NN)rtorch.nnr rModulerrr r?s  -"))-r