:jddlmZddlmZddlmZmZddlmZm Z ddl m Z ddl mZddlmZd d lmZmZmZd d lmZd d lmZmZmZd d lmZmZmZgdZGddej@Z!Gddej@Z"dede#de!fdZ$GddeZ%eede%jLfdejNfdddejNddee%d e(dee#d!eed"ede!f d#Z)y)$) OrderedDict)partial)AnyOptional)nnTensor) functional)SemanticSegmentation)_log_api_usage_once)register_modelWeights WeightsEnum)_VOC_CATEGORIES)_ovewrite_value_paramhandle_legacy_interfaceIntermediateLayerGetter)mobilenet_v3_largeMobileNet_V3_Large_Weights MobileNetV3)LRASPP!LRASPP_MobileNet_V3_Large_Weightslraspp_mobilenet_v3_largec neZdZdZ d dej dededededdf fd Zd ede e effd Z xZ S) ra Implements a Lite R-ASPP Network for semantic segmentation from `"Searching for MobileNetV3" `_. 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 "high" for the high level feature map and "low" for the low level feature map. low_channels (int): the number of channels of the low level features. high_channels (int): the number of channels of the high level features. num_classes (int, optional): number of output classes of the model (including the background). inter_channels (int, optional): the number of channels for intermediate computations. backbone low_channels high_channels num_classesinter_channelsreturnNclt|t|||_t |||||_y)N)super__init__r r LRASPPHead classifier)selfrrrrr __class__s {/home/mcse/projects/srt_converter/srt-converter-venv/lib/python3.12/site-packages/torchvision/models/segmentation/lraspp.pyr$zLRASPP.__init__#s1 D!  $\=+~^inputc|j|}|j|}tj||jdddd}t }||d<|S)NbilinearFsizemode align_cornersout)rr&F interpolateshaper)r'r+featuresr3results r)forwardzLRASPP.forward+sS=='ooh'mmCekk"#&6ZW\]u  r*)) __name__ __module__ __qualname____doc__rModuleintr$rdictstrr9 __classcell__r(s@r)rrsk  sv_ _14_EH_WZ_lo_ _VS&[(9r*rc NeZdZdededededdf fd Zdeeefdefd ZxZ S) r%rrrr r!Nc t|tjtj||ddtj |tj d|_tjtjdtj||ddtj|_ tj||d|_ tj||d|_ y)NF)biasT)inplace) r#r$r SequentialConv2d BatchNorm2dReLUcbrAdaptiveAvgPool2dSigmoidscalelow_classifierhigh_classifier)r'rrrr r(s r)r$zLRASPPHead.__init__7s == IIm^QU C NN> * GGD !  ]]   # IIm^QU C JJL  !ii k1E!yyaHr*r+c|d}|d}|j|}|j|}||z}tj||jdddd}|j ||j |zS)Nlowhighr-r.Fr/)rNrQr4r5r6rRrS)r'r+rUrVxss r)r9zLRASPPHead.forwardFsxElV} HHTN JJt  E MM!#))BC.zQV W""3'$*>*>q*AAAr*) r;r<r=r@r$rArBrr9rCrDs@r)r%r%6sR IS I I3 I`c Ihl I BT#v+. B6 Br*r%rrr!c X|j}dgt|Dcgc]\}}t|dds|c}}zt|dz gz}|d}|d}||j}||j}t |t |dt |di }t||||Scc}}w) Nr_is_cnFrGrUrV) return_layers)r7 enumerategetattrlen out_channelsrrBr) rrib stage_indiceslow_poshigh_posrrs r)_lraspp_mobilenetv3rgRs  HC8)<\A8UZ@[1\\`cdl`mpq`q_rrMBGR HG$11LX&33M&xG eUXYaUbdj?klH (L- EE]s B&B&c ReZdZedeeddeddddd d id d d dZeZy)rzJhttps://download.pytorch.org/models/lraspp_mobilenet_v3_large-d234d4ea.pthi) resize_sizei"(1)rGrGz]https://github.com/pytorch/vision/tree/main/references/segmentation#lraspp_mobilenet_v3_largezCOCO-val2017-VOC-labelsg33333L@gV@)miou pixel_accg㥛 @g{G(@z These weights were trained on a subset of COCO, using only the 20 categories that are present in the Pascal VOC dataset. ) num_params categoriesmin_sizerecipe_metrics_ops _file_size_docs)url transformsmetaN) r;r<r=rrr rCOCO_WITH_VOC_LABELS_V1DEFAULTr*r)rr`sS% X/SA!)u) !%,  ,&Gr*r pretrainedpretrained_backbone)weightsweights_backboneNT)r|progressrr}r|r~r}kwargsc f|jddr tdtj|}t j|}|&d}t d|t |jd}n|d}t|d }t||}|"|j|j|d |S) a|Constructs a Lite R-ASPP Network model with a MobileNetV3-Large backbone from `Searching for MobileNetV3 `_ paper. .. betastatus:: segmentation module Args: weights (:class:`~torchvision.models.segmentation.LRASPP_MobileNet_V3_Large_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.segmentation.LRASPP_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: parameters passed to the ``torchvision.models.segmentation.LRASPP`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.segmentation.LRASPP_MobileNet_V3_Large_Weights :members: aux_lossFz&This model does not use auxiliary lossNrrmT)r|dilated)r~ check_hash) popNotImplementedErrorrverifyrrr`rvrrgload_state_dictget_state_dict)r|r~rr}rrmodels r)rrzsLzz*e$!"JKK/66w?G1889IJ+M;GLLYeLfHgh   !*:DIH + 6E g44hSW4XY Lr*)* collectionsr functoolsrtypingrrtorchrrtorch.nnr r4transforms._presetsr utilsr _apirrr_metar_utilsrrr mobilenetv3rrr__all__r?rr%r@rgrrw IMAGENET1K_V1boolrryr*r)rs%# $7(77#\\UU W RYY FBB8 F+ FC FF F& &4 <TT U+-G-U-UV <@!%=W=e=e 3 7 833# 3 9: 3  3 3  3r*