o ,h3@sXddlZddlmZddlmZddlmZmZmZm Z ddl Z ddl m Z m Z ddl mZddlmZdd lmZd d lmZd d lmZmZmZd d lmZd dlmZmZd dlmZm Z ddlmZ!ddl"m#Z#ddl$m%Z%ddl&m'Z'm(Z(ddgZ)de*de*de*dede j+fde j,f ddZ-de*de*dede j+fde j,fdd Z.d!e j+fd"d#Z/Gd$d%d%e j+Z0Gd&d'd'e(Z1Gd(d)d)e(Z2Gd*d+d+e j+Z3d,e ej4ej5fd-e*dede j+ffd.d/Z6Gd0ddeZ7eed1e7j8fd2e j9fd3dd4de j9ddd5d6ee7d7e:d8ee*d9ee d:ee*deede j+fd;ede'fdc s^eZdZdeedeedededejfffdd Zdee d e e e ffd d Z Z S) SSDLiteHeadr num_anchors num_classesr#.cs,tt|||||_t||||_dSN)super__init__SSDLiteClassificationHeadclassification_headSSDLiteRegressionHeadregression_head)selfr r@rAr# __class__r*r+rDRs zSSDLiteHead.__init__xr$cCs||||dS)N)bbox_regression cls_logits)rHrF)rIrLr*r*r+forwardYszSSDLiteHead.forward)__name__ __module__ __qualname__listintrrModulerDr dictstrrO __classcell__r*r*rJr+r?Qs &r?c s@eZdZdeedeedededejfffdd ZZ S)rEr r@rAr#.csNt}t||D]\}}|t|||d|q t|t||dS)Nr r ModuleListzipappendr,r>rCrD)rIr r@rAr#rNchannelsanchorsrJr*r+rDas z"SSDLiteClassificationHead.__init__ rPrQrRrSrTrrrUrDrXr*r*rJr+rE`s rEcs<eZdZdeedeededejfffdd ZZ S)rGr r@r#.csNt}t||D]\}}|t|d|d|q t|t|ddS)Nr rY)rIr r@r#bbox_regr]r^rJr*r+rDls zSSDLiteRegressionHead.__init__r_r*r*rJr+rGks4rGc s^eZdZ  ddejdededejfdedef fd d Zd e d e e e ffd dZ Z S) SSDLiteFeatureExtractorMobileNet?backbonec4_posr#. width_mult min_depthc stt|||jrtdttjg|d|||jdRtj||jddg||ddR|_fdd}t t |dj |d|t |d|d|t |d|d|t |d|d |g}t |||_ dS) Nz0backbone[c4_pos].use_res_connect should be Falserrcstt|SrB)maxrT)drhrgr*r+sz;SSDLiteFeatureExtractorMobileNet.__init__..i)rCrDr use_res_connect ValueErrorrr'blockfeaturesrZr0r!r>extra)rIrerfr#rgrh get_depthrtrJrkr+rDus$  $* z)SSDLiteFeatureExtractorMobileNet.__init__rLr$cCsVg}|jD] }||}||q|jD] }||}||qtddt|DS)NcSsg|] \}}t||fqSr*)rW).0ivr*r*r+ sz.)rsr\rtr enumerate)rIrLoutputrrr*r*r+rOs    z(SSDLiteFeatureExtractorMobileNet.forward)rcrd)rPrQrRrrUrTrfloatrDr rVrWrOrXr*r*rJr+rbts "!rbretrainable_layerscCs|j}dgddt|Dt|dg}t|}d|kr'|ks,tdtd|dkr4t|n|||}|d|D]}|D]}|dqFq@t||d|S)NrcSs g|] \}}t|ddr|qS)_is_cnF)getattr)rvrwbr*r*r+rys z(_mobilenet_extractor..rzYtrainable_layers should be in the range [0, {num_stages}], instead got {trainable_layers}F)rsrzlenrq parametersrequires_grad_rb)rer}r# stage_indices num_stages freeze_beforer parameterr*r*r+_mobilenet_extractors&  rc @s8eZdZedededddddiidd d d d ZeZd S)rzShttps://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-val2017box_mapgL5@g-?gt*@zSThese weights were produced by following a similar training recipe as on the paper.) num_params categoriesmin_sizerecipe_metrics_ops _file_size_docs)url transformsmetaN)rPrQrRrr rCOCO_V1DEFAULTr*r*r*r+rs$ pretrainedpretrained_backbone)weightsweights_backboneT)rprogressrArtrainable_backbone_layersr#rrrArrkwargscKst|}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.NrAr[gMbP?r2)epsmomentum)rrr# reduced_tail)@rcSsg|]}ddgqS)rr r*)rv_r*r*r+ry,sz1ssdlite320_mobilenet_v3_large..g?gffffff?) min_ratio 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) score_thresh nms_threshdetections_per_imgtopk_candidates image_mean image_stdheadT)r check_hashr*)rverifyrwarningswarnrrrrrr BatchNorm2drr>rrrange det_utilsretrieve_out_channelsnum_anchors_per_location aspect_ratiosrqrr?load_state_dictget_state_dict)rrrArrr#r reduce_tailreranchor_generatorr!r@defaultsmodelr*r*r+rsp 8      )r?rErGrb MobileNetV2 MobileNetV3rrr IMAGENET1K_V1boolrr*r*r*r+s            &  0