a hN<@sddlmZddlmZmZmZddlZddlmZddlm Z ddl m Z ddl m Z dd lmZmZmZdd lmZdd lmZmZgd Ze ee d ddZGdddejZGdddejZeeeeeedddZdeddZGdddeZ GdddeZ!GdddeZ"Gd d!d!eZ#eed"e j$fd#dd$d%ee eeed&d'd(Z%eed"e!j$fd#dd$d%ee!eeed&d)d*Z&eed"e"j$fd#dd$d%ee"eeed&d+d,Z'eed"e#j$fd#dd$d%ee#eeed&d-d.Z(dS)/)partial)AnyCallableOptionalN)Tensor)ImageClassification)_log_api_usage_once)register_modelWeights WeightsEnum)_IMAGENET_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface) ShuffleNetV2ShuffleNet_V2_X0_5_WeightsShuffleNet_V2_X1_0_WeightsShuffleNet_V2_X1_5_WeightsShuffleNet_V2_X2_0_Weightsshufflenet_v2_x0_5shufflenet_v2_x1_0shufflenet_v2_x1_5shufflenet_v2_x2_0)xgroupsreturncCsP|\}}}}||}||||||}t|dd}|||||}|S)Nr r)sizeviewtorchZ transpose contiguous)rrZ batchsizeZ num_channelsheightwidthZchannels_per_groupr#M/var/www/auris/lib/python3.9/site-packages/torchvision/models/shufflenetv2.pychannel_shuffles r%c sZeZdZeeeddfdd Zedeeeeeeejdd d Z e e d d d Z Z S)InvertedResidualN)inpoupstridercsNtd|krdks(ntd||_|d}|jdkrh||d>krhtd|d|d|d|jdkrt|j||d|jdd t|tj||ddd d d t|tj d d|_ n t|_ ttj|jdkr|n||ddd d d t|tj d d|j||d|jdd t|tj||ddd d d t|tj d d|_ dS)Nr zillegal stride valuerzInvalid combination of stride z, inp z and oup zB values. If stride == 1 then inp should be equal to oup // 2 << 1. kernel_sizer)paddingrF)r,r)r-biasTZinplace) super__init__ ValueErrorr)nn Sequentialdepthwise_conv BatchNorm2dConv2dReLUbranch1branch2)selfr'r(r)Zbranch_features __class__r#r$r1,sF      zInvertedResidual.__init__r rF)ior,r)r-r.rc Cstj|||||||dS)N)r.r)r3r7)r>r?r,r)r-r.r#r#r$r5VszInvertedResidual.depthwise_convrrcCsb|jdkr6|jddd\}}tj|||fdd}ntj||||fdd}t|d}|S)Nr r)Zdim)r)chunkrcatr:r9r%)r;rx1Zx2outr#r#r$forward\s   zInvertedResidual.forward)r rF) __name__ __module__ __qualname__intr1 staticmethodboolr3r7r5rrE __classcell__r#r#r<r$r&+s* r&csbeZdZdefeeeeeedejfddfdd Z e e ddd Z e e dd d Z Z S) ri.N)stages_repeatsstages_out_channels num_classesinverted_residualrc sdtt|t|dkr&tdt|dkr:td||_d}|jd}ttj||ddddd t |tj d d |_ |}tj dddd |_ |||d ddD}t|||jddD]R\}} }|||dg} t| dD]} | |||dqt||tj| |}q|jd}ttj||ddddd t |tj d d |_t|||_dS)Nr*z2expected stages_repeats as list of 3 positive intsz7expected stages_out_channels as list of 5 positive intsrrr F)r.Tr/r+cSsg|]}d|qS)Zstager#).0r>r#r#r$ z)ShuffleNetV2.__init__..)rr*)r0r1r lenr2Z_stage_out_channelsr3r4r7r6r8conv1Z MaxPool2dmaxpoolziprangeappendsetattrconv5ZLinearfc) r;rMrNrOrPZinput_channelsZoutput_channelsZ stage_namesnameZrepeatsseqr>r<r#r$r1isB        zShuffleNetV2.__init__r@cCsX||}||}||}||}||}||}|ddg}||}|S)Nrr*)rXrYZstage2Zstage3Zstage4r^meanr_r;rr#r#r$ _forward_impls       zShuffleNetV2._forward_implcCs ||S)N)rdrcr#r#r$rEszShuffleNetV2.forward)rFrGrHr&listrIrr3Moduler1rrdrErLr#r#r<r$rhs 0 r)weightsprogressargskwargsrcOsL|durt|dt|jdt|i|}|durH||j|dd|S)NrO categoriesT)rhZ check_hash)rrWmetarZload_state_dictZget_state_dict)rgrhrirjmodelr#r#r$ _shufflenetv2s rn)r r z2https://github.com/ericsun99/Shufflenet-v2-Pytorch)Zmin_sizerkrecipec @sDeZdZedeeddiedddddid d d d d ZeZdS)rzDhttps://download.pytorch.org/models/shufflenetv2_x0.5-f707e7126e.pth crop_sizei ImageNet-1Kg-FN@g9voT@zacc@1zacc@5g{Gz?gT㥛 @VThese weights were trained from scratch to reproduce closely the results of the paper. num_params_metrics_ops _file_size_docsurlZ transformsrlN rFrGrHr rr _COMMON_META IMAGENET1K_V1DEFAULTr#r#r#r$rs$ rc @sDeZdZedeeddiedddddid d d d d ZeZdS)rzBhttps://download.pytorch.org/models/shufflenetv2_x1-5666bf0f80.pthrprqi"rsgI +WQ@gNbX9V@rtg(\?gE!@rurvr|Nr~r#r#r#r$rs$ rc @sHeZdZedeedddieddddd d id d d ddZeZdS)rzBhttps://download.pytorch.org/models/shufflenetv2_x1_5-3c479a10.pthrprrZ resize_size+https://github.com/pytorch/vision/pull/5906iv5rsg9v?R@g/$V@rtgl?gw/+@ These weights were trained from scratch by using TorchVision's `new training recipe `_. rorwrxryrzr{r|Nr~r#r#r#r$rs& rc @sHeZdZedeedddieddddd d id d d ddZeZdS)rzBhttps://download.pytorch.org/models/shufflenetv2_x2_0-8be3c8ee.pthrprrriprsgQS@gMb@W@rtg-?g+n<@rrr|Nr~r#r#r#r$rs& rZ pretrained)rgT)rgrh)rgrhrjrcKs(t|}t||gdgdfi|S)a  Constructs a ShuffleNetV2 architecture with 0.5x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. Args: weights (:class:`~torchvision.models.ShuffleNet_V2_X0_5_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ShuffleNet_V2_X0_5_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.shufflenetv2.ShuffleNetV2`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ShuffleNet_V2_X0_5_Weights :members: rUrU)0`)rverifyrnrgrhrjr#r#r$rs rcKs(t|}t||gdgdfi|S)a  Constructs a ShuffleNetV2 architecture with 1.0x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. Args: weights (:class:`~torchvision.models.ShuffleNet_V2_X1_0_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ShuffleNet_V2_X1_0_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.shufflenetv2.ShuffleNetV2`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ShuffleNet_V2_X1_0_Weights :members: r)rtrir)rrrnrr#r#r$r>s rcKs(t|}t||gdgdfi|S)a  Constructs a ShuffleNetV2 architecture with 1.5x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. Args: weights (:class:`~torchvision.models.ShuffleNet_V2_X1_5_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ShuffleNet_V2_X1_5_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.shufflenetv2.ShuffleNetV2`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ShuffleNet_V2_X1_5_Weights :members: r)ri`ir)rrrnrr#r#r$r]s rcKs(t|}t||gdgdfi|S)a  Constructs a ShuffleNetV2 architecture with 2.0x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. Args: weights (:class:`~torchvision.models.ShuffleNet_V2_X2_0_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.ShuffleNet_V2_X2_0_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.shufflenetv2.ShuffleNetV2`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.ShuffleNet_V2_X2_0_Weights :members: r)riii)rrrnrr#r#r$r|s r)) functoolsrtypingrrrrZtorch.nnr3rZtransforms._presetsrutilsr Z_apir r r Z_metar_utilsrr__all__rIr%rfr&rrKrnrrrrrrrrrrr#r#r#r$sn       =B