U (d: @sddlmZddlmZmZmZddlZddlmZddlm Z ddl m Z dd l mZmZdd lmZdd lmZmZmZdd lmZmZmZdd lmZmZmZmZmZddlm Z ddl!m"Z"dddddddgZ#Gddde Z$Gdddej%Z&Gdddej%Z'Gdddej%Z(Gdd d ej)Z*ee+ee,e$d!d"d#Z-ed$d%d&Z.Gd'ddeZ/Gd(ddeZ0Gd)ddeZ1ee+ee,e$d!d*d+Z2ed,e/j3fd-ej4fd.dd/ddej4d0ee/e,ee+ee,eeee$d1d2dZ5ed,e0j3fd-ej4fd.dd/ddej4d0ee0e,ee+ee,eeee$d1d3dZ6ed,e1j3fd-ej4fd.dd/ddej4d0ee1e,ee+ee,eeee$d1d4dZ7dd5lm8Z8e8e/j3j9e0j3j9e1j3j9d6Z:dS)7)partial)AnyListOptionalN)nn) functional)SemanticSegmentation) WeightsEnumWeights)_VOC_CATEGORIES)IntermediateLayerGetterhandle_legacy_interface_ovewrite_value_param) MobileNetV3MobileNet_V3_Large_Weightsmobilenet_v3_large)ResNetresnet50 resnet101ResNet50_WeightsResNet101_Weights)_SimpleSegmentationModel)FCNHead DeepLabV3DeepLabV3_ResNet50_WeightsDeepLabV3_ResNet101_Weights$DeepLabV3_MobileNet_V3_Large_Weightsdeeplabv3_mobilenet_v3_largedeeplabv3_resnet50deeplabv3_resnet101c@seZdZdZdS)ra Implements DeepLabV3 model from `"Rethinking Atrous Convolution for Semantic Image Segmentation" `_. Args: backbone (nn.Module): the network used to compute the features for the model. The backbone should return an OrderedDict[Tensor], with the key being "out" for the last feature map used, and "aux" if an auxiliary classifier is used. classifier (nn.Module): module that takes the "out" element returned from the backbone and returns a dense prediction. aux_classifier (nn.Module, optional): auxiliary classifier used during training N)__name__ __module__ __qualname____doc__r'r'M/tmp/pip-unpacked-wheel-vx7f76es/torchvision/models/segmentation/deeplabv3.pyrscs&eZdZeeddfdd ZZS) DeepLabHeadN) in_channels num_classesreturnc sHtt|dddgtjddddddtdttd|ddS) N $rrF)paddingbias)super__init__ASPPrConv2d BatchNorm2dReLU)selfr*r+ __class__r'r(r41s zDeepLabHead.__init__r#r$r%intr4 __classcell__r'r'r:r(r)0sr)cs(eZdZeeeddfdd ZZS)ASPPConvN)r* out_channelsdilationr,cs6tj||d||ddt|tg}tj|dS)NrF)r1rAr2)rr6r7r8r3r4)r9r*r@rAmodulesr:r'r(r4<s zASPPConv.__init__r<r'r'r:r(r?;sr?cs:eZdZeeddfdd ZejejdddZZS) ASPPPoolingN)r*r@r,c s4ttdtj||dddt|tdS)NrFr2)r3r4rZAdaptiveAvgPool2dr6r7r8)r9r*r@r:r'r(r4Fs zASPPPooling.__init__xr,cCs2|jdd}|D] }||}qtj||dddS)NZbilinearF)sizemodeZ align_corners)shapeFZ interpolate)r9rFrHmodr'r'r(forwardNs zASPPPooling.forward) r#r$r%r=r4torchTensorrMr>r'r'r:r(rCEsrCcsBeZdZd eeeeddfdd ZejejdddZZ S) r5r0N)r* atrous_ratesr@r,c stg}|ttj||dddt|tt|}|D]}|t |||qF|t ||t ||_ ttjt |j ||dddt|ttd|_dS)NrFrDg?)r3r4appendr Sequentialr6r7r8tupler?rCZ ModuleListconvslenZDropoutproject)r9r*rPr@rBZratesZrater:r'r(r4Vs  $ z ASPP.__init__rEcCs6g}|jD]}|||q tj|dd}||S)Nr)Zdim)rTrQrNcatrV)r9rFZ_resconvresr'r'r(rMls  z ASPP.forward)r0) r#r$r%r=rr4rNrOrMr>r'r'r:r(r5Usr5)backboner+auxr,cCsHddi}|rd|d<t||d}|r.td|nd}td|}t|||S)NZlayer4outr[Zlayer3 return_layersii)rrr)r)rZr+r[r^aux_classifier classifierr'r'r(_deeplabv3_resnetts  ra)rrz These weights were trained on a subset of COCO, using only the 20 categories that are present in the Pascal VOC dataset. ) categoriesZmin_sizeZ_docsc @s<eZdZedeeddedddddd id d ZeZd S) rzHhttps://download.pytorch.org/models/deeplabv3_resnet50_coco-cd0a2569.pthZ resize_sizeijzVhttps://github.com/pytorch/vision/tree/main/references/segmentation#deeplabv3_resnet50COCO-val2017-VOC-labelsgP@皙W@ZmiouZ pixel_accZ num_paramsZrecipeZ_metricsurlZ transformsmetaN r#r$r%r rr _COMMON_METACOCO_WITH_VOC_LABELS_V1DEFAULTr'r'r'r(rs c @s<eZdZedeeddedddddd id d ZeZd S) rzIhttps://download.pytorch.org/models/deeplabv3_resnet101_coco-586e9e4e.pthrcrdijzQhttps://github.com/pytorch/vision/tree/main/references/segmentation#fcn_resnet101regP@rfrgrhriNrlr'r'r'r(rs c @s<eZdZedeeddedddddd id d ZeZd S) rzMhttps://download.pytorch.org/models/deeplabv3_mobilenet_v3_large-fc3c493d.pthrcrdiPKz`https://github.com/pytorch/vision/tree/main/references/segmentation#deeplabv3_mobilenet_v3_largeregfffff&N@gV@rgrhriNrlr'r'r'r(rs c Cs|j}dgddt|Dt|dg}|d}||j}|d}||j}t|di}|rld|t|<t||d }|rt||nd} t||} t|| | S) NrcSs g|]\}}t|ddr|qS)Z_is_cnF)getattr).0ibr'r'r( s z*_deeplabv3_mobilenetv3..rr\r[r]) features enumeraterUr@strrrr)r) rZr+r[Z stage_indicesZout_posZ out_inplanesZaux_posZ aux_inplanesr^r_r`r'r'r(_deeplabv3_mobilenetv3s&      rzZ pretrainedZpretrained_backbone)weightsweights_backboneT)r{progressr+aux_lossr|)r{r}r+r~r|kwargsr,cKst|}t|}|dk r@d}t|t|jd}t|d}n |dkrLd}t|dddgd}t|||}|dk r||j |d|S)adConstructs a DeepLabV3 model with a ResNet-50 backbone. .. betastatus:: segmentation module Reference: `Rethinking Atrous Convolution for Semantic Image Segmentation `__. Args: weights (:class:`~torchvision.models.segmentation.DeepLabV3_ResNet50_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.segmentation.DeepLabV3_ResNet50_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. num_classes (int, optional): number of output classes of the model (including the background) aux_loss (bool, optional): If True, it uses an auxiliary loss weights_backbone (:class:`~torchvision.models.ResNet50_Weights`, optional): The pretrained weights for the backbone **kwargs: unused .. autoclass:: torchvision.models.segmentation.DeepLabV3_ResNet50_Weights :members: NrbTFr{Zreplace_stride_with_dilationr}) rverifyrrrUrkrraload_state_dictget_state_dictr{r}r+r~r|rrZmodelr'r'r(r!s$    cKst|}t|}|dk r@d}t|t|jd}t|d}n |dkrLd}t|dddgd}t|||}|dk r||j |d|S)aiConstructs a DeepLabV3 model with a ResNet-101 backbone. .. betastatus:: segmentation module Reference: `Rethinking Atrous Convolution for Semantic Image Segmentation `__. Args: weights (:class:`~torchvision.models.segmentation.DeepLabV3_ResNet101_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.segmentation.DeepLabV3_ResNet101_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. num_classes (int, optional): number of output classes of the model (including the background) aux_loss (bool, optional): If True, it uses an auxiliary loss weights_backbone (:class:`~torchvision.models.ResNet101_Weights`, optional): The pretrained weights for the backbone **kwargs: unused .. autoclass:: torchvision.models.segmentation.DeepLabV3_ResNet101_Weights :members: NrbTrFrr) rrrrrUrkrrarrrr'r'r(r"s$    cKst|}t|}|dk r@d}t|t|jd}t|d}n |dkrLd}t|dd}t|||}|dk r~||j |d|S)akConstructs a DeepLabV3 model with a MobileNetV3-Large backbone. Reference: `Rethinking Atrous Convolution for Semantic Image Segmentation `__. Args: weights (:class:`~torchvision.models.segmentation.DeepLabV3_MobileNet_V3_Large_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.segmentation.DeepLabV3_MobileNet_V3_Large_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. num_classes (int, optional): number of output classes of the model (including the background) aux_loss (bool, optional): If True, it uses an auxiliary loss weights_backbone (:class:`~torchvision.models.MobileNet_V3_Large_Weights`, optional): The pretrained weights for the backbone **kwargs: unused .. autoclass:: torchvision.models.segmentation.DeepLabV3_MobileNet_V3_Large_Weights :members: NrbTr)r{Zdilatedr) rrrrrUrkrrzrrrr'r'r(r Ks"     ) _ModelURLs)Zdeeplabv3_resnet50_cocoZdeeplabv3_resnet101_cocoZ!deeplabv3_mobilenet_v3_large_coco); functoolsrtypingrrrrNrZtorch.nnrrKZtransforms._presetsr Z_apir r Z_metar _utilsrrrZ mobilenetv3rrrZresnetrrrrrrZfcnr__all__rrRr)r?rCModuler5r=boolrarmrrrrzrnZ IMAGENET1K_V1r!r"r rrjZ model_urlsr'r'r'r(s              332