a h?@sddlmZddlmZddlmZmZmZddlZddlm Z m Z ddl m Z m ZddlmZdd lmZd d lmZmZmZd d lmZd d lmZmZmZgdZGdddZGddde jZ Gddde jZ!d,e"e#e$e$edddZ%e&ee'eee$ee!dddZ(dedZ)Gdd d eZ*Gd!d"d"eZ+eed#e*j,fd$dd%d&ee*e$ee!d'd(d)Z-eed#e+j,fd$dd%d&ee+e$ee!d'd*d+Z.dS)-)Sequence)partial)AnyCallableOptionalN)nnTensor)Conv2dNormActivationSqueezeExcitation)ImageClassification)_log_api_usage_once)register_modelWeights WeightsEnum)_IMAGENET_CATEGORIES)_make_divisible_ovewrite_named_paramhandle_legacy_interface) MobileNetV3MobileNet_V3_Large_WeightsMobileNet_V3_Small_Weightsmobilenet_v3_largemobilenet_v3_smallc @s>eZdZeeeeeeeeed ddZeeedddZ dS)InvertedResidualConfig) input_channelskernelexpanded_channels out_channelsuse_se activationstridedilation width_multc CsP||| |_||_||| |_||| |_||_|dk|_||_||_dS)NHS) adjust_channelsrrrrr use_hsr"r#) selfrrrrr r!r"r#r$r)L/var/www/auris/lib/python3.9/site-packages/torchvision/models/mobilenetv3.py__init__s  zInvertedResidualConfig.__init__Zchannelsr$cCst||dS)N)rr,r)r)r*r&0sz&InvertedResidualConfig.adjust_channelsN) __name__ __module__ __qualname__intboolstrfloatr+ staticmethodr&r)r)r)r*rs rcsXeZdZeeejdfeedej fedej fdfdd Z e e dddZ Z S) InvertedResidual)Zscale_activation.)cnf norm_layerse_layerc s$td|jkr dks*ntd|jdko>|j|jk|_g}|jrRtj ntj }|j |jkr| t |j|j d||d|jdkrdn|j}| t |j |j |j||j|j ||d|jrt|j dd}| ||j || t |j |jd|ddtj||_|j|_|jdk|_dS)Nrr zillegal stride value kernel_sizer8activation_layer)r;r"r#groupsr8r<r-)superr+r" ValueErrorrruse_res_connectr'r HardswishZReLUrappendr r#rr r SequentialblockZ_is_cn)r(r7r8r9layersr<r"Zsqueeze_channels __class__r)r*r+7sP     zInvertedResidual.__init__)inputreturncCs||}|jr||7}|SN)rErA)r(rIresultr)r)r*forwardos zInvertedResidual.forward)r.r/r0rSElayerrZ HardsigmoidrrModuler+rrM __classcell__r)r)rGr*r65s   8r6c sreZdZd eeeeeedej feedej fe e ddfdd Z e e dd d Ze e dd d ZZS)rN皙?.)inverted_residual_setting last_channel num_classesrEr8dropoutkwargsrJc stt||s tdn$t|trz(MobileNetV3.__init__..zDThe inverted_residual_setting should be List[InvertedResidualConfig]NgMbP?g{Gz?)ZepsZmomentumrr )r;r"r8r<rr:T)inplace)pr`Zfan_out)mode)"r?r+r r@rXrall TypeErrorr6rrZ BatchNorm2drrCr rBrrDfeaturesZAdaptiveAvgPool2davgpoolZLinearZDropout classifiermodulesZConv2dinitZkaiming_normal_ZweightZbiasZzeros_Z GroupNormZones_Znormal_)r(rSrTrUrEr8rVrWrFZfirstconv_output_channelsr7Zlastconv_input_channelsZlastconv_output_channelsmrGr)r*r+wsr             zMobileNetV3.__init__)xrJcCs.||}||}t|d}||}|S)Nr)rerftorchflattenrgr(rkr)r)r* _forward_impls     zMobileNetV3._forward_implcCs ||SrK)rornr)r)r*rMszMobileNetV3.forward)rQNNrR)r.r/r0listrr1rrrrOr4rr+rrorMrPr)r)rGr*rvs [ r?F)archr$ reduced_taildilatedrWc Ks|rdnd}|rdnd}tt|d}ttj|d}|dkr|dddddddd|ddd d dddd|d dd d dddd|d d d d dddd|d d dd dddd|d d dd dddd|d ddddddd|dddddddd|dddddddd|dddddddd|dddddddd|dddddddd|dd dd|ddd||d|d d|d|ddd||d|d d|d|ddd|g} |d|} n"|dkr|dddddddd|ddd d dddd|d ddd dddd|d d dd dddd|d d dd dddd|d d dd dddd|d d dddddd|dd dddddd|dd d d|ddd||d|d d!|d|ddd||d|d d!|d|ddd|g } |d"|} ntd#|| | fS)$Nr r)r$rr]FZRE@H(TxPr%ipiiirX`0i i@izUnsupported model type )rrr&r@) rrr$rsrtrWZreduce_dividerr#Z bneck_confr&rSrTr)r)r*_mobilenet_v3_confsL          r)rSrTweightsprogressrWrJcKsP|durt|dt|jdt||fi|}|durL||j|dd|S)NrU categoriesT)rZ check_hash)rlenmetarZload_state_dictZget_state_dict)rSrTrrrWmodelr)r)r* _mobilenet_v3s r)rr)Zmin_sizerc @s~eZdZedeeddiedddddd id d d d dZedeedddiedddddd id ddd dZeZ dS)rzChttps://download.pytorch.org/models/mobilenet_v3_large-8738ca79.pth crop_sizeiS^https://github.com/pytorch/vision/tree/main/references/classification#mobilenetv3-large--small ImageNet-1Kg R@g(\V@zacc@1zacc@5g-?gw/5@zJThese weights were trained from scratch by using a simple training recipe.Z num_paramsZrecipeZ_metricsZ_ops _file_sizeZ_docsurlZ transformsrzChttps://download.pytorch.org/models/mobilenet_v3_large-5c1a4163.pth)rZ resize_sizezHhttps://github.com/pytorch/vision/issues/3995#new-recipe-with-reg-tuninggK7R@gNbX9$W@gZd5@a/ These weights improve marginally upon the results of the original paper by using a modified version of TorchVision's `new training recipe `_. N) r.r/r0rrr _COMMON_META IMAGENET1K_V1Z IMAGENET1K_V2DEFAULTr)r)r)r*r)sJ  rc @sFeZdZedeeddiedddddd id d d d dZeZdS)rzChttps://download.pytorch.org/models/mobilenet_v3_small-047dcff4.pthrri&rrgnP@g}?5^U@rgv/?g r#@z} These weights improve upon the results of the original paper by using a simple training recipe. rrN) r.r/r0rrr rrrr)r)r)r*rUs& rZ pretrained)rT)rr)rrrWrJcKs2t|}tdi|\}}t||||fi|S)a Constructs a large MobileNetV3 architecture from `Searching for MobileNetV3 `__. Args: weights (:class:`~torchvision.models.MobileNet_V3_Large_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.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. **kwargs: parameters passed to the ``torchvision.models.mobilenet.MobileNetV3`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.MobileNet_V3_Large_Weights :members: r)r)rverifyrrrrrWrSrTr)r)r*rms rcKs2t|}tdi|\}}t||||fi|S)a Constructs a small MobileNetV3 architecture from `Searching for MobileNetV3 `__. Args: weights (:class:`~torchvision.models.MobileNet_V3_Small_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.MobileNet_V3_Small_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. **kwargs: parameters passed to the ``torchvision.models.mobilenet.MobileNetV3`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.MobileNet_V3_Small_Weights :members: r)r)rrrrrr)r)r*rs r)rqFF)/collections.abcr functoolsrtypingrrrrlrrZops.miscr r rNZtransforms._presetsr utilsr Z_apirrrZ_metar_utilsrrr__all__rrOr6rr3r4r2rrpr1rrrrrrrr)r)r)r*sX      Ak 2 ,