o eiK7@sdZddlZddlmZddlmZddlmZddlmZddl m Z dd l m Z d d l mZGd d d ejZGdddejZGdddejZGdddejZGdddejZe Gddde Ze ddGdddeZddgZdS)zrPyTorch UperNet model. Based on OpenMMLab's implementation, found in https://github.com/open-mmlab/mmsegmentation.N)nn)CrossEntropyLoss) load_backbone)SemanticSegmenterOutput)PreTrainedModel)auto_docstring) UperNetConfigcseZdZdZ   ddededeeeefBdeeeefBeBd ed eeeefBd d ffd d Zde j d e j fddZ Z S)UperNetConvModulez A convolutional block that bundles conv/norm/activation layers. This block simplifies the usage of convolution layers, which are commonly used with a norm layer (e.g., BatchNorm) and activation layer (e.g., ReLU). rFr in_channels out_channels kernel_sizepaddingbiasdilationreturnNcs<ttj||||||d|_t||_t|_dS)N)r r rrrr) super__init__rConv2dconv BatchNorm2d batch_normReLU activation)selfr r rrrr __class__j/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/upernet/modeling_upernet.pyr!s  zUperNetConvModule.__init__inputcCs"||}||}||}|SN)rrr)rr outputrrrforward6s   zUperNetConvModule.forward)rFr ) __name__ __module__ __qualname____doc__inttuplestrboolrtorchTensorr# __classcell__rrrrr s* r csDeZdZdedededdffdd Zdejdejfd d ZZS) UperNetPyramidPoolingBlock pool_scaler channelsrNcsLtt|t||ddg|_t|jD] \}}|t||qdS)Nr r) rrrAdaptiveAvgPool2dr layers enumerate add_moduler*)rr0r r1ilayerrrrr?s  z#UperNetPyramidPoolingBlock.__init__r cCs|}|jD]}||}q|Sr!)r4)rr hidden_stater8rrrr#Hs  z"UperNetPyramidPoolingBlock.forward) r$r%r&r(rr,r-r#r.rrrrr/>s r/c sXeZdZdZdeedfdedededdf fd d Zd ej de ej fd d Z Z S)UperNetPyramidPoolingModulea} Pyramid Pooling Module (PPM) used in PSPNet. Args: pool_scales (`tuple[int]`): Pooling scales used in Pooling Pyramid Module. in_channels (`int`): Input channels. channels (`int`): Channels after modules, before conv_seg. align_corners (`bool`): align_corners argument of F.interpolate. pool_scales.r r1 align_cornersrNcsht||_||_||_||_g|_t|D]\}}t|||d}|j || t ||qdS)N)r0r r1) rrr;r<r r1blocksr5r/appendr6r*)rr;r r1r<r7r0blockrrrr^s  z$UperNetPyramidPoolingModule.__init__xcCsHg}|jD]}||}tjj||ddd|jd}||q|S)Nbilinearsizemoder<)r=r functional interpolaterDr<r>)rr@ppm_outsppmppm_outupsampled_ppm_outrrrr#js  z#UperNetPyramidPoolingModule.forward) r$r%r&r'r)r(r+rr,r-listr#r.rrrrr:Os*" r:cs>eZdZdZfddZddZdejdejfdd ZZ S) UperNetHeadz Unified Perceptual Parsing for Scene Understanding. This head is the implementation of [UPerNet](https://huggingface.co/papers/1807.10221). cst||_|j|_||_|j|_d|_tj |j|j dd|_ t |j|jd|j|jd|_ t|jdt|j|j|jddd|_t|_t|_|jddD] }t||jdd}t|j|jddd}|j||j|qVtt|j|j|jddd|_dS)NFr r2)r<rrr)rrconfigr;r hidden_sizer1r<rr num_labels classifierr: psp_modulesr len bottleneck ModuleList lateral_convs fpn_convsr>fpn_bottleneck)rrPr l_convfpn_convrrrr{s@     zUperNetHead.__init__cCs:|d}|g}|||tj|dd}||}|S)NrNr dim)extendrTr,catrV)rinputsr@psp_outsr"rrr psp_forwards  zUperNetHead.psp_forwardencoder_hidden_statesrcsfddtjDt}t|dddD]$}|djdd}|dtjj ||dj d|d<q fd dt|dD}|dt|dddD]}tjj |||djdddj d||<qbt j |dd } |}|}|S) Ncsg|] \}}||qSrr).0r7 lateral_conv)rdrr sz'UperNetHead.forward..r rrNrArBrCcsg|] }j||qSr)rY)rer7)lateralsrrrrgsr])r5rXr>rcrUrangeshaperrFrGr<r,r`rZrS)rrdused_backbone_levelsr7 prev_shapefpn_outsr"r)rdrhrrr#s$   zUperNetHead.forward) r$r%r&r'rrcr,r-r#r.rrrrrMus  ' rMc sXeZdZdZ ddededeeeefBdd ffd d Zd ejdejfd dZ Z S)UperNetFCNHeada Fully Convolution Networks for Semantic Segmentation. This head is the implementation of [FCNNet](https://huggingface.co/papers/1411.4038>). Args: config: Configuration. in_channels (int): Number of input channels. kernel_size (int): The kernel size for convs in the head. Default: 3. dilation (int): The dilation rate for convs in the head. Default: 1. rArr in_indexrrrNc st||_|jdur||n|j|_|j|_|j|_|j |_ ||_ |d|}g}| t |j|j|||dt|jdD]}| t |j|j|||dqA|jdkr]t|_ntj||_|j rvt |j|j|j||dd|_tj|j|jdd|_dS)NrA)rrrr rrOr2)rrrPauxiliary_in_channelsr auxiliary_channelsr1auxiliary_num_convs num_convsauxiliary_concat_input concat_inputror>r rirIdentityconvs Sequentialconv_catrrRrS) rrPr rorr conv_paddingrwr7rrrrs:     zUperNetFCNHead.__init__rdcCs@||j}||}|jr|tj||gdd}||}|S)Nr r])rorwruryr,r`rS)rrd hidden_statesr"rrrr#s   zUperNetFCNHead.forward)rArr ) r$r%r&r'r(r)rr,r-r#r.rrrrrns&rnc@s"eZdZUeed<dZdZgZdS)UperNetPreTrainedModelrP pixel_values)imageN)r$r%r&r __annotations__main_input_nameinput_modalities_no_split_modulesrrrrr| s r|zW UperNet framework leveraging any vision backbone e.g. for ADE20k, CityScapes. ) custom_introcsjeZdZfddZe     d dejdBdedBdedBdejdBdedBd ee Bf d d Z Z S) UperNetForSemanticSegmentationcsPt|t||_t||jjd|_|jrt||jjdnd|_ | dS)N)r ) rrrbackbonerMr1 decode_headuse_auxiliary_headrnauxiliary_head post_init)rrPrrrrs   z'UperNetForSemanticSegmentation.__init__Nr}output_attentionsoutput_hidden_stateslabels return_dictrcKsl|dur|jjdkrtd|dur|n|jj}|dur|n|jj}|dur(|n|jj}|jj|||d}|j}| |} t j j | |j ddddd} d} |jdurg||} t j j | |j ddddd} d} |durt|jjd } | | |} | dur| | |} | |jj| 7} |s|r| f|dd}n | f|dd}| dur| f|S|St| | |j|jd S) a labels (`torch.LongTensor` of shape `(batch_size, height, width)`, *optional*): Ground truth semantic segmentation maps for computing the loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels > 1`, a classification loss is computed (Cross-Entropy). Examples: ```python >>> from transformers import AutoImageProcessor, UperNetForSemanticSegmentation >>> from PIL import Image >>> from huggingface_hub import hf_hub_download >>> image_processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-convnext-tiny") >>> model = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-convnext-tiny") >>> filepath = hf_hub_download( ... repo_id="hf-internal-testing/fixtures_ade20k", filename="ADE_val_00000001.jpg", repo_type="dataset" ... ) >>> image = Image.open(filepath).convert("RGB") >>> inputs = image_processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> logits = outputs.logits # shape (batch_size, num_labels, height, width) >>> list(logits.shape) [1, 150, 512, 512] ```Nr z/The number of labels should be greater than one)rrrArBFrC) ignore_index)losslogitsr{ attentions)rPrR ValueErroruse_return_dictrrrforward_with_filtered_kwargs feature_mapsrrrFrGrjrrloss_ignore_indexauxiliary_loss_weightrr{r)rr}rrrrkwargsoutputsfeaturesrauxiliary_logitsrloss_fctauxiliary_lossr"rrrr#%sH%     z&UperNetForSemanticSegmentation.forward)NNNNN) r$r%r&rrr,r-r+r)rr#r.rrrrrs* r)r'r,rtorch.nnrbackbone_utilsrmodeling_outputsrmodeling_utilsrutilsrconfiguration_upernetr Moduler r/r:rMrnr|r__all__rrrrs*        #&T@ d