[TU]5 [R"UUSUUSS9Ul[R "X1R S9Ul[R"5Ul g)NrF) in_channels out_channels kernel_sizestridepaddingr2)eps) super__init__rr,convr-batch_norm_eps batch_normReLUrelu)r4r%r?r@rBrC __class__s r!rFVitMatteBasicConv3x3.__init__HsU II#%   ..;P;PQGGI r clURU5nURU5nURU5nU$N)rGrIrKr4 hidden_states r!forwardVitMatteBasicConv3x3.forwardUs2yy. |4 yy. r )rIrGrK)r rrrrrrFrRr __classcell__rLs@r!r<r<Cs r r<c2^\rSrSrSrU4SjrSrSrU=r$)VitMatteConvStream]z[ Simple ConvStream containing a series of basic conv3x3 layers to extract detail features. c>[TU]5 SnURbURRnURn[ R "5UlU/U-Ul[[UR5S- 5HInURUnURUS-nURR[XU55 MK g)Nr ) rErFbackbone_config num_channelsconvstream_hidden_sizesr ModuleListconvs conv_chansrangelenappendr<)r4r%r?r@iin_chan_ out_chan_rLs r!rFVitMatteConvStream.__init__bs   ! ! - 00==K55 ]]_ &-,6s4??+a/0Aq)HA.I JJ  26YO P1r cSU0nUn[[UR55H-nURU"U5nS[US-5-nX2U'M/ U$)Ndetailed_feature_map_0detailed_feature_map_r )rcrdrastr)r4r&out_dict embeddingsrfname_s r!rRVitMatteConvStream.forwardusZ,l;! s4::'AAz2J+c!a%j8E(UO( r )rbrarUrWs@r!rYrY]sQ&r rYc2^\rSrSrSrU4SjrSrSrU=r$)VitMatteFusionBlockzT Simple fusion block to fuse features from ConvStream and Plain Vision Transformer. cD>[TU]5 [XUSSS9Ulg)Nr )rBrC)rErFr<rG)r4r%r?r@rLs r!rFVitMatteFusionBlock.__init__s" (lST^_` r c[RRUSSSS9n[R"X#/SS9nUR U5nU$)NrTbilinearF) scale_factormode align_cornersr )dim)r functional interpolatercatrG)r4featuresdetailed_feature_mapupscaled_featuresouts r!rRVitMatteFusionBlock.forwardsHMM55hQU_ot5uii-AqIiin r )rGrUrWs@r!rsrssar rsc2^\rSrSrSrU4SjrSrSrU=r$) VitMatteHeadzB Simple Matting Head, containing only conv3x3 and conv1x1 layers. c &>[TU]5 URSnSn[R"[R "X#SSSS9[R "U5[R"S5[R "USSSSS95Ulg)Nrr )rArBrCTr) rErFfusion_hidden_sizesr Sequentialr,r-rJ matting_convs)r4r%r? mid_channelsrLs r!rFVitMatteHead.__init__sr 004  ]] IIkQqRS T NN< ( GGDM IIlA1Q J  r c(URU5nU$rOrrPs r!rRVitMatteHead.forwards)),7 r rrUrWs@r!rrs  r rc2^\rSrSrSrU4SjrSrSrU=r$)VitMatteDetailCaptureModulez? Simple and lightweight Detail Capture Module for ViT Matting. c >[TU]5 [UR5[UR5S-:wa [ S5eXl[U5UlURRUl [R"5Ul UR/UR-Ul[[UR5S- 5HXnURR![#UURUURUS-*-URUS-S95 MZ [%U5Ulg)Nr z_The length of fusion_hidden_sizes should be equal to the length of convstream_hidden_sizes + 1.)r%r?r@)rErFrdrr_ ValueErrorr%rY convstreamrbrr` fusion_blocks hidden_sizefusion_channelsrcrersr matting_head)r4r%rfrLs r!rF$VitMatteDetailCaptureModule.__init__s  v)) *c&2P2P.QTU.U Uq  ,V4//44]]_ & 2 23f6P6PPs4//0145A    % %#! $ 4 4Q 7$//APQE(:S S!%!5!5a!e!< 6)0r c:URU5n[[UR55HBnS[ [UR5U- S- 5-nURU"XU5nMD [ R "URU55nU$)Nrlr )rrcrdrrmrsigmoidr)r4rr&detail_featuresrfdetailed_feature_map_namers r!rR#VitMatteDetailCaptureModule.forwards//,7s4--./A(?#c$J\J\F]`aFadeFeBf(f %))!,XG`7abH0t00:; r )r%rbrrrrrUrWs@r!rrs12r rzX ViTMatte framework leveraging any vision backbone e.g. for ADE20k, CityScapes. c ^\rSrSrU4Sjr\S S\\RS\\ S\\ S\\RS\\ 4 Sjj5r S r U=r $) VitMatteForImageMattingc>[TU]U5 Xl[U5Ul[ U5UlUR5 grO)rErFr%r backbonerdecoder post_init)r4r%rLs r!rF VitMatteForImageMatting.__init__s9   %f- 26:  r r&output_attentionsoutput_hidden_stateslabels return_dictcUbUOURRnUbUOURRnUbUOURRnSnUb [ S5eUR R XUS9nURSnURX5n U(dU 4USS-n UbU4U -$U $[UU URURS9$)ap labels (`torch.LongTensor` of shape `(batch_size, height, width)`, *optional*): Ground truth image matting for computing the loss. Examples: ```python >>> from transformers import VitMatteImageProcessor, VitMatteForImageMatting >>> import torch >>> from PIL import Image >>> from huggingface_hub import hf_hub_download >>> processor = VitMatteImageProcessor.from_pretrained("hustvl/vitmatte-small-composition-1k") >>> model = VitMatteForImageMatting.from_pretrained("hustvl/vitmatte-small-composition-1k") >>> filepath = hf_hub_download( ... repo_id="hf-internal-testing/image-matting-fixtures", filename="image.png", repo_type="dataset" ... ) >>> image = Image.open(filepath).convert("RGB") >>> filepath = hf_hub_download( ... repo_id="hf-internal-testing/image-matting-fixtures", filename="trimap.png", repo_type="dataset" ... ) >>> trimap = Image.open(filepath).convert("L") >>> # prepare image + trimap for the model >>> inputs = processor(images=image, trimaps=trimap, return_tensors="pt") >>> with torch.no_grad(): ... alphas = model(**inputs).alphas >>> print(alphas.shape) torch.Size([1, 1, 640, 960]) ```NzTraining is not yet supported)rrrr )rrrr) r%use_return_dictrrNotImplementedErrorrforward_with_filtered_kwargs feature_mapsrrrr) r4r&rrrrroutputsrroutputs r!rRVitMatteForImageMatting.forwardsR&1%B 'tn B & B d^ B B r r)r dataclassesrtypingrrrmodeling_utilsrutilsrr utils.backbone_utilsr configuration_vitmatter rr#r:r<rYrsrrr__all__rr r!rs! -012  ::  :$ )o ) )2994  F"))"2990&"))&R N 5N  N b %&? @r