jCddlmZddlmZmZmZmZddlZddlm Z ddlm Z ddl m Z ddl mZdd lmZmZmZdd lmZdd lmZmZdd l mZmZmZmZd dlmZmZmZgdZ Gdde j!Z"Gdde j#Z$dee%dee%deede&de&dede$fdZ'dedddd Z(Gd!d"eZ)Gd#d$eZ*Gd%d&eZ+Gd'd(eZ,ed)*ed+d,f-dd.d/d0deee)efde&de&dede$f d1Z-ed2*ed+d3f-dd.d/d0deee*efde&de&dede$f d4Z.ed5*ed+d6f-dd.d/d0deee+efde&de&dede$f d7Z/ed8*ed+d9f-dd.d/d0deee,efde&de&dede$f d:Z0dS);)partial)AnyListOptionalUnionN)Tensor) shufflenetv2)ImageClassification)register_modelWeights WeightsEnum)_IMAGENET_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface)ShuffleNet_V2_X0_5_WeightsShuffleNet_V2_X1_0_WeightsShuffleNet_V2_X1_5_WeightsShuffleNet_V2_X2_0_Weights) _fuse_modules _replace_reluquantize_model) QuantizableShuffleNetV2#ShuffleNet_V2_X0_5_QuantizedWeights#ShuffleNet_V2_X1_0_QuantizedWeights#ShuffleNet_V2_X1_5_QuantizedWeights#ShuffleNet_V2_X2_0_QuantizedWeightsshufflenet_v2_x0_5shufflenet_v2_x1_0shufflenet_v2_x1_5shufflenet_v2_x2_0c<eZdZdededdffd ZdedefdZxZS)QuantizableInvertedResidualargskwargsreturnNctj|i|tj|_dSN)super__init__nn quantizedFloatFunctionalcatselfr&r' __class__s ]C:\Projects\Historical\venv\Lib\site-packages\torchvision/models/quantization/shufflenetv2.pyr,z$QuantizableInvertedResidual.__init__$s8$)&)))<//11xcf|jdkrL|dd\}}|j|||gd}nD|j||||gd}t j|d}|S)Nrr )dim)stridechunkr0branch2branch1r channel_shuffle)r2r6x1x2outs r4forwardz#QuantizableInvertedResidual.forward(s ;!  WWQAW&&FB(,,DLL$4$451,==CC(,, QaAq,IIC*322 r5)__name__ __module__ __qualname__rr,rrA __classcell__r3s@r4r%r%#st2c2S2T222222  F        r5r%cZeZdZdededdffd ZdedefdZd deeddfd Z xZ S) rr&r'r(Nctj|dti|tjj|_tjj|_ dS)Ninverted_residual) r+r,r%torchao quantization QuantStubquant DeQuantStubdequantr1s r4r,z QuantizableShuffleNetV2.__init__6sZ$X2MXQWXXXX*4466 x,88:: r5r6c||}||}||}|Sr*)rN _forward_implrP)r2r6s r4rAzQuantizableShuffleNetV2.forward;s8 JJqMM   q ! ! LLOOr5is_qatc|jD]!\}}|dvr|t|gdg|d"|D]}t |t urnt |jjdkrt|jddggd g|dt|jgdd d ggd g|ddS) aBFuse conv/bn/relu modules in shufflenetv2 model Fuse conv+bn+relu/ conv+relu/conv+bn modules to prepare for quantization. Model is modified in place. .. note:: Note that this operation does not change numerics and the model after modification is in floating point )conv1conv5N)012T)inplacerrWrX)rY34r[r\)567) _modulesitemsrmodulestyper%lenr<r;)r2rSnamems r4 fuse_modelz"QuantizableShuffleNetV2.fuse_modelAs }**,, J JGD!)))ama///!2FDIIII  AAww555qy)//1122Q66!!)sCj///-JF\`aaaaI$__sCj///B   r5r*) rBrCrDrr,rrArboolrgrErFs@r4rr4s;c;S;T;;;;;; F $4r5rstages_repeatsstages_out_channelsweightsprogressquantizer'r(c |Nt|dt|jdd|jvrt|d|jd|dd}t ||fi|}t ||rt |||*|||d|S)N num_classes categoriesbackendfbgemmT)rl check_hash) rrdmetapoprrrload_state_dictget_state_dict)rirjrkrlrmr'rqmodels r4 _shufflenetv2ryZsfmSl9S5T5TUUU  $ $ !&)W\)5L M M MjjH--G #N4G R R6 R RE%'ug&&& g44hSW4XXYYY Lr5)rrrrzdhttps://github.com/pytorch/vision/tree/main/references/classification#post-training-quantized-modelsz These weights were produced by doing Post Training Quantization (eager mode) on top of the unquantized weights listed below. )min_sizerprqrecipe_docsc leZdZedeediedejddddid d d  Z e Z d S)rzShttps://download.pytorch.org/models/quantized/shufflenetv2_x0.5_fbgemm-00845098.pth crop_sizei ImageNet-1Kg#~jL@gRS@zacc@1zacc@5g{Gz?gjt? num_params unquantized_metrics_ops _file_sizeurl transformsrtN) rBrCrDrrr _COMMON_METAr IMAGENET1K_V1IMAGENET1K_FBGEMM_V1DEFAULTr5r4rrs"7 a7.#>>>    !5C##       "#GGGr5rc leZdZedeediedejddddid d d  Z e Z d S)rzQhttps://download.pytorch.org/models/quantized/shufflenetv2_x1_fbgemm-1e62bb32.pthr~ri"rgףp= Q@gh|?U@rg(\?gy&1@rrN) rBrCrDrrr rrrrrrr5r4rrs"7 _7.#>>>    !5C##       "#GGGr5rc peZdZedeeddieddejddd d id d d Z e Z dS)rzShttps://download.pytorch.org/models/quantized/shufflenetv2_x1_5_fbgemm-d7401f05.pthr~r resize_size+https://github.com/pytorch/vision/pull/5906iv5rgSR@g̬V@rgl?gK7A` @r{rrrrrrN) rBrCrDrrr rrrrrrr5r4rr"7 a7.#3OOO    C!5C##       $#GGGr5rc peZdZedeeddieddejddd d id d d Z e Z dS)rzShttps://download.pytorch.org/models/quantized/shufflenetv2_x2_0_fbgemm-5cac526c.pthr~rrriprg-R@gZd;W@rg-?g|?5@rrN) rBrCrDrrr rrrrrrr5r4rrrr5rquantized_shufflenet_v2_x0_5)re pretrainedc^|ddr tjn tjSNrmF)getrrrrr's r4r, ::j% ( (6:OO ' 5r5)rkTFrkrlrmc t|rtnt|}tgdgdf|||d|S)aQ Constructs a ShuffleNetV2 with 0.5x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. .. note:: Note that ``quantize = True`` returns a quantized model with 8 bit weights. Quantized models only support inference and run on CPUs. GPU inference is not yet supported. Args: weights (:class:`~torchvision.models.quantization.ShuffleNet_V2_X0_5_QuantizedWeights` or :class:`~torchvision.models.ShuffleNet_V2_X0_5_Weights`, optional): The pretrained weights for the model. See :class:`~torchvision.models.quantization.ShuffleNet_V2_X0_5_QuantizedWeights` 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. quantize (bool, optional): If True, return a quantized version of the model. Default is False. **kwargs: parameters passed to the ``torchvision.models.quantization.ShuffleNet_V2_X0_5_QuantizedWeights`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.quantization.ShuffleNet_V2_X0_5_QuantizedWeights :members: .. autoclass:: torchvision.models.ShuffleNet_V2_X0_5_Weights :members: :noindex: r)0`r)rrverifyryrkrlrmr's r4r r sa`7?^22D^ffgnooG  *** 4;hYa  ek  r5quantized_shufflenet_v2_x1_0c^|ddr tjn tjSr)rrrrrrs r4rrrr5c t|rtnt|}tgdgdf|||d|S)aQ Constructs a ShuffleNetV2 with 1.0x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. .. note:: Note that ``quantize = True`` returns a quantized model with 8 bit weights. Quantized models only support inference and run on CPUs. GPU inference is not yet supported. Args: weights (:class:`~torchvision.models.quantization.ShuffleNet_V2_X1_0_QuantizedWeights` or :class:`~torchvision.models.ShuffleNet_V2_X1_0_Weights`, optional): The pretrained weights for the model. See :class:`~torchvision.models.quantization.ShuffleNet_V2_X1_0_QuantizedWeights` 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. quantize (bool, optional): If True, return a quantized version of the model. Default is False. **kwargs: parameters passed to the ``torchvision.models.quantization.ShuffleNet_V2_X1_0_QuantizedWeights`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.quantization.ShuffleNet_V2_X1_0_QuantizedWeights :members: .. autoclass:: torchvision.models.ShuffleNet_V2_X1_0_Weights :members: :noindex: r)rtrirr)rrrryrs r4r!r! a`7?^22D^ffgnooG  ,,, 6=[c  gm  r5quantized_shufflenet_v2_x1_5c^|ddr tjn tjSr)rrrrrrs r4rrFrr5c t|rtnt|}tgdgdf|||d|S)aQ Constructs a ShuffleNetV2 with 1.5x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. .. note:: Note that ``quantize = True`` returns a quantized model with 8 bit weights. Quantized models only support inference and run on CPUs. GPU inference is not yet supported. Args: weights (:class:`~torchvision.models.quantization.ShuffleNet_V2_X1_5_QuantizedWeights` or :class:`~torchvision.models.ShuffleNet_V2_X1_5_Weights`, optional): The pretrained weights for the model. See :class:`~torchvision.models.quantization.ShuffleNet_V2_X1_5_QuantizedWeights` 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. quantize (bool, optional): If True, return a quantized version of the model. Default is False. **kwargs: parameters passed to the ``torchvision.models.quantization.ShuffleNet_V2_X1_5_QuantizedWeights`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.quantization.ShuffleNet_V2_X1_5_QuantizedWeights :members: .. autoclass:: torchvision.models.ShuffleNet_V2_X1_5_Weights :members: :noindex: r)ri`irr)rrrryrs r4r"r"Brr5quantized_shufflenet_v2_x2_0c^|ddr tjn tjSr)rrrrrrs r4rr|rr5c t|rtnt|}tgdgdf|||d|S)aQ Constructs a ShuffleNetV2 with 2.0x output channels, as described in `ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design `__. .. note:: Note that ``quantize = True`` returns a quantized model with 8 bit weights. Quantized models only support inference and run on CPUs. GPU inference is not yet supported. Args: weights (:class:`~torchvision.models.quantization.ShuffleNet_V2_X2_0_QuantizedWeights` or :class:`~torchvision.models.ShuffleNet_V2_X2_0_Weights`, optional): The pretrained weights for the model. See :class:`~torchvision.models.quantization.ShuffleNet_V2_X2_0_QuantizedWeights` 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. quantize (bool, optional): If True, return a quantized version of the model. Default is False. **kwargs: parameters passed to the ``torchvision.models.quantization.ShuffleNet_V2_X2_0_QuantizedWeights`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.quantization.ShuffleNet_V2_X2_0_QuantizedWeights :members: .. autoclass:: torchvision.models.ShuffleNet_V2_X2_0_Weights :members: :noindex: r)riiir)rrrryrs r4r#r#xrr5)1 functoolsrtypingrrrrrJtorch.nnr-rtorchvision.modelsr transforms._presetsr _apir rr_metar_utilsrrrrrrutilsrrr__all__InvertedResidualr% ShuffleNetV2rintrhryrrrrrr r!r"r#rr5r4rs------------ ++++++6666667777777777((((((CCCCCCCC @?????????   ,"?"#####l7###LIck "     6&t   #####+###*#####+###*#####+###,#####+###,3444 6 6 ae *** e?A[[\ ]** *  *  ***54*Z3444 6 6 ae *** e?A[[\ ]** *  *  ***54*Z3444 6 6 ae *** e?A[[\ ]** *  *  ***54*Z3444 6 6 ae *** e?A[[\ ]** *  *  ***54***r5