U ʹj3 @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$m%Z%ddl&m'Z'ddl(m)Z)m*Z*ddgZ+e,e,e,ede j-fe j.dddZ/e,e,ede j-fe j.dddZ0e j-ddd Z1Gd!d"d"e j-Z2Gd#d$d$e*Z3Gd%d&d&e*Z4Gd'd(d(e j-Z5e ej6ej7fe,ede j-fd)d*d+Z8Gd,ddeZ9eed-e9j:fd.e"j;fd/dd0de"j;ddd1e e9er?)rDr"r;r<r%rIchannelsanchorsrEr,r-r?as z"SSDLiteClassificationHead.__init__ rKrLrMrrNrr rOr?rQr,r,rEr-r@`s  r@cs8eZdZeeeeedejfdfdd ZZ S)rB.)r"r;r%csNt}t||D] \}}|t|d|d|qt|t|ddS)Nr rR)rDr"r;r%Zbbox_regrVrWrEr,r-r?ls zSSDLiteRegressionHead.__init__rXr,r,rEr-rBksrBcsPeZdZd ejeedejfeedfdd Ze e e e fddd Z Z S) SSDLiteFeatureExtractorMobileNet?.)backbonec4_posr% width_mult min_depthc stt|||jr$tdttj|d|||jdftj||jddf||dd|_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=)maxrN)dr`r_r,r-z;SSDLiteFeatureExtractorMobileNet.__init__..i)r>r?rZuse_res_connect ValueErrorr r)blockfeaturesrSr/r#r8extra)rDr]r^r%r_r`Z get_depthrlrErcr-r?us$   (z)SSDLiteFeatureExtractorMobileNet.__init__rGcCsVg}|jD]}||}||q |jD]}||}||q(tddt|DS)NcSsg|]\}}t||fqSr,)rP).0ivr,r,r- sz.)rkrUrlr enumerate)rDrHoutputrjr,r,r-rJs    z(SSDLiteFeatureExtractorMobileNet.forward)r[r\)rKrLrMr rOrNrfloatr?r rrPrJrQr,r,rEr-rZts !rZ)r]trainable_layersr%cCs|j}dgddt|Dt|dg}t|}d|krH|ksRntd|dkrbt|n |||}|d|D]}|D]}|dqqzt||d|S)NrcSs g|]\}}t|ddr|qS)Z_is_cnF)getattr)rmrnbr,r,r-rps z(_mobilenet_extractor..rzYtrainable_layers should be in the range [0, {num_stages}], instead got {trainable_layers}F)rkrqlenri parametersZrequires_grad_rZ)r]rtr%Z stage_indicesZ num_stagesZ freeze_beforervZ parameterr,r,r-_mobilenet_extractors& rzc @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)rKrLrMrrrCOCO_V1DEFAULTr,r,r,r-r s$Z pretrainedZpretrained_backbone)weightsweights_backboneT)rprogressr<rtrainable_backbone_layersr%)rrr<rrr%kwargsr&cKst|}t|}d|kr&td|dk rJd}td|t|jd}n |dkrVd}t|dk pf|dk |dd}|dk}|dkrt t j dd d }t f||||d |}|dkrt |t|||}d } td dtdDddd} t|| } | } t| t| jkr*tdt| dt| jdddddddgdddgd} | |}t|| | |fdt| | ||i|}|dk r||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<r{[gMbP?r1)epsZmomentum)rrr%Z reduced_tail)@rcSsg|] }ddgqS)rr r,)rm_r,r,r-rp,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,g?)Z score_threshZ nms_threshZdetections_per_imgZtopk_candidatesZ image_meanZ image_stdheadT)rZ check_hash)r verifyrwarningswarnrrxr~rrr Z BatchNorm2drr8rzrrange det_utilsZretrieve_out_channelsZnum_anchors_per_locationZ aspect_ratiosrirr9Zload_state_dictZget_state_dict)rrr<rrr%rZ reduce_tailr]rZanchor_generatorr#r;defaultsmodelr,r,r-r!s|8       )>r collectionsr functoolsrtypingrrrrrr r4r r Zops.miscr Ztransforms._presetsrutilsrrZ_apirrr_metarrrrZ mobilenetv3rrrZ anchor_utilsrZbackbone_utilsrZssdrr__all__rNrOr)r.r/r8r9r@rBrZZ MobileNetV2Z MobileNetV3rzr rZ IMAGENET1K_V1boolr!r,r,r,r-st                1