o )i3@s`ddlZddlmZddlmZddlmZmZmZm Z m Z m Z ddl Z ddl m Z mZddlmZddlmZdd lmZd d lmZd d lmZmZmZd d lmZd dlmZmZd dl m!Z!m"Z"ddlmZ#ddl$m%Z%ddl&m'Z'ddl(m)Z)m*Z*ddgZ+de,de,de,dede j-fde j.f ddZ/de,de,dede j-fde j.fdd Z0d!e j-fd"d#Z1Gd$d%d%e j-Z2Gd&d'd'e*Z3Gd(d)d)e*Z4Gd*d+d+e j-Z5d,e ej6ej7fd-e,dede j-ffd.d/Z8Gd0ddeZ9eed1e9j:fd2e"j;fd3dd4de"j;ddd5d6e e9d7er?)rDr"r;r<r%rHchannelsanchorsrEr,r-r?as z"SSDLiteClassificationHead.__init__ rJrKrLrrMrr rNr?rPr,r,rEr-r@`s r@cs<eZdZdeedeededejfffdd ZZ S)rBr"r;r%.csNt}t||D]\}}|t|d|d|q t|t|ddS)Nr rQ)rDr"r;r%Zbbox_regrUrVrEr,r-r?ls zSSDLiteRegressionHead.__init__rWr,r,rEr-rBks4rBc s^eZdZ  ddejdededejfdedef fd d Zd e d e e e ffd dZ Z S) SSDLiteFeatureExtractorMobileNet?backbonec4_posr%. width_mult min_depthc stt|||jrtdttjg|d|||jdRtj||jddg||ddR|_fdd}t t |dj |d|t |d|d|t |d|d|t |d|d |g}t |||_ dS) Nz0backbone[c4_pos].use_res_connect should be Falserrcstt|Sr=)maxrM)dr_r^r,r-sz;SSDLiteFeatureExtractorMobileNet.__init__..i)r>r?rZuse_res_connect ValueErrorr r)blockfeaturesrRr/r#r9extra)rDr\r]r%r^r_Z get_depthrjrErbr-r?us$  $* z)SSDLiteFeatureExtractorMobileNet.__init__rGr&cCsVg}|jD] }||}||q|jD] }||}||qtddt|DS)NcSsg|] \}}t||fqSr,)rO).0ivr,r,r- sz.)rirTrjr enumerate)rDrGoutputrhr,r,r-rIs    z(SSDLiteFeatureExtractorMobileNet.forward)rZr[)rJrKrLr rNrMrfloatr?r rrOrIrPr,r,rEr-rYts "!rYr\trainable_layerscCs|j}dgddt|Dt|dg}t|}d|kr'|ks,tdtd|dkr4t|n|||}|d|D]}|D]}|dqFq@t||d|S)NrcSs g|] \}}t|ddr|qS)Z_is_cnF)getattr)rkrlbr,r,r-rns z(_mobilenet_extractor..rzYtrainable_layers should be in the range [0, {num_stages}], instead got {trainable_layers}F)rirolenrg parametersZrequires_grad_rY)r\rrr%Z stage_indicesZ num_stagesZ freeze_beforertZ parameterr,r,r-_mobilenet_extractors&  rxc @s8eZdZedededddddiidd d d d ZeZd S)r zShttps://download.pytorch.org/models/ssdlite320_mobilenet_v3_large_coco-a79551df.pthi}4)rrz]https://github.com/pytorch/vision/tree/main/references/detection#ssdlite320-mobilenetv3-largez COCO-val2017Zbox_mapgL5@g-?gt*@zSThese weights were produced by following a similar training recipe as on the paper.)Z num_params categoriesZmin_sizeZrecipeZ_metricsZ_ops _file_sizeZ_docs)urlZ transformsmetaN)rJrKrLrrrCOCO_V1DEFAULTr,r,r,r-r s$Z pretrainedZpretrained_backbone)weightsweights_backboneT)rprogressr<rtrainable_backbone_layersr%rrr<rrkwargscKst|}t|}d|vrtd|dur%d}td|t|jd}n|dur+d}t|dup3|du|dd}|du}|durHt t j dd d }t d||||d |}|dur\t |t|||}d } td dtdDddd} t|| } | } t| t| jkrtdt| dt| jddddgdgdd} i| |}t|| | |fdt| | ||i|}|dur||j|dd|S)aSSDlite model architecture with input size 320x320 and a MobileNetV3 Large backbone, as described at `Searching for MobileNetV3 `__ and `MobileNetV2: Inverted Residuals and Linear Bottlenecks `__. .. betastatus:: detection module See :func:`~torchvision.models.detection.ssd300_vgg16` for more details. Example: >>> model = torchvision.models.detection.ssdlite320_mobilenet_v3_large(weights=SSDLite320_MobileNet_V3_Large_Weights.DEFAULT) >>> model.eval() >>> x = [torch.rand(3, 320, 320), torch.rand(3, 500, 400)] >>> predictions = model(x) Args: weights (:class:`~torchvision.models.detection.SSDLite320_MobileNet_V3_Large_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.detection.SSDLite320_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). weights_backbone (:class:`~torchvision.models.MobileNet_V3_Large_Weights`, optional): The pretrained weights for the backbone. trainable_backbone_layers (int, optional): number of trainable (not frozen) layers starting from final block. Valid values are between 0 and 6, with 6 meaning all backbone layers are trainable. If ``None`` is passed (the default) this value is set to 6. norm_layer (callable, optional): Module specifying the normalization layer to use. **kwargs: parameters passed to the ``torchvision.models.detection.ssd.SSD`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.detection.SSDLite320_MobileNet_V3_Large_Weights :members: sizez?The size of the model is already fixed; ignoring the parameter.Nr<ry[gMbP?r1)ZepsZmomentum)rrr%Z reduced_tail)@rcSsg|]}ddgqS)rr r,)rk_r,r,r-rn,sz1ssdlite320_mobilenet_v3_large..g?gffffff?)Z min_ratioZ max_ratioz4The length of the output channels from the backbone z? do not match the length of the anchor generator aspect ratios g?i,)?rr)Z score_threshZ nms_threshZdetections_per_imgZtopk_candidatesZ image_meanZ image_stdheadT)rZ check_hashr,)r verifyrwarningswarnrrvr|rrr Z BatchNorm2drr9rxrrange det_utilsZretrieve_out_channelsZnum_anchors_per_locationZ aspect_ratiosrgrr:Zload_state_dictZget_state_dict)rrr<rrr%rZ reduce_tailr\rZanchor_generatorr#r;defaultsmodelr,r,r-r!sp 8      )>r collectionsr functoolsrtypingrrrrrr r5r r Zops.miscr Ztransforms._presetsrutilsrrZ_apirrr_metarrrrZ mobilenetv3rrrZ anchor_utilsrZbackbone_utilsrZssdrr__all__rMrNr)r.r/r9r:r@rBrYZ MobileNetV2Z MobileNetV3rxr r}Z IMAGENET1K_V1boolr!r,r,r,r-s             &  0