o hA@sddlmZddlmZddlmZmZmZmZddl m Z ddl m Z ddl mZddlmZd d lmZmZmZd d lmZd d lmZmZgd ZGddde jZGddde jZGddde jZGddde j Z!Gddde j Z"Gddde jZ#Gddde jZ$Gddde j Z%de&ee!e"fdee&eeeefd e'e(d!ed"e j fd#eed$e)d%ed&e%fd'd(Z*d)ed*d+d,Z+Gd-d.d.eZ,Gd/d0d0eZ-Gd1d2d2eZ.eed3e,j/fd4dd5d6d#ee,d$e)d%ed&e%fd7d8Z0eed3e-j/fd4dd5d6d#ee-d$e)d%ed&e%fd9d:Z1eed3e.j/fd4dd5d6d#ee.d$e)d%ed&e%fd;d<Z2d d=lm3Z3e3e,j/j4e-j/j4e.j/j4d>Z5dS)?)Sequence)partial)AnyCallableOptionalUnionN)Tensor)VideoClassification)_log_api_usage_once)register_modelWeights WeightsEnum)_KINETICS400_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface) VideoResNetR3D_18_WeightsMC3_18_WeightsR2Plus1D_18_Weightsr3d_18mc3_18 r2plus1d_18c ^eZdZ d dededeedededdf fd d Zededeeeeffd d ZZ S) Conv3DSimpleN in_planes out_planes midplanesstridepaddingreturncstj||d||dddS)N)r r r FZ in_channelsZ out_channels kernel_sizer r!biassuper__init__selfrrrr r! __class__e/home/www/facesmatcher.com/frenv_anti/lib/python3.10/site-packages/torchvision/models/video/resnet.pyr(s zConv3DSimple.__init__cC |||fSNr-r r-r-r.get_downsample_stride( z"Conv3DSimple.get_downsample_strideNrr __name__ __module__ __qualname__intrr( staticmethodtupler2 __classcell__r-r-r+r.r" &rc sXeZdZd dedededededdf fd d Zededeeeeffd d ZZS) Conv2Plus1Drrrrr r!r"Nc s`ttj||dd||fd||fddt|tjddtj||d|ddf|ddfdddS) Nrr r rrFr$r r!r%TZinplacer rrr'r(nnConv3d BatchNorm3dReLUr)r+r-r.r(.s zConv2Plus1D.__init__cCr/r0r-r1r-r-r.r2?r3z!Conv2Plus1D.get_downsample_striderr) r6r7r8r9r(r:r;r2r<r-r-r+r.r>-s(&r>c r)Conv3DNoTemporalNrrrrr r!r"cs(tj||dd||fd||fdddS)Nr?rrFr#r&r)r+r-r.r(Es zConv3DNoTemporal.__init__cCs d||fSNrr-r1r-r-r.r2Rr3z&Conv3DNoTemporal.get_downsample_strider4r5r-r-r+r.rIDr=rIc sbeZdZdZ  ddedededejfdedeejd df fd d Z d e d e fd dZ Z S) BasicBlockrNinplanesplanes conv_builder.r downsampler"cs||ddd|ddd|}tt|||||t|tjdd|_t||||t||_tjdd|_||_ ||_ dS)Nr TrA) r'r(rD SequentialrFrGconv1conv2relurOr r*rLrMrNr rOrr+r-r.r([s(  zBasicBlock.__init__xcCsB|}||}||}|jdur||}||7}||}|Sr0)rQrRrOrSr*rUZresidualoutr-r-r.forwardns     zBasicBlock.forwardrN r6r7r8 expansionr9rrDModulerr(rrXr<r-r-r+r.rKWs$ rKc sbeZdZdZ  ddedededejfded eejd df fd d Z d e d e fddZ Z S) BottleneckrNrLrMrN.r rOr"cst||ddd|ddd|}ttj||dddt|tjdd|_t|||||t|tjdd|_ttj|||j dddt||j |_ tjdd|_ ||_ ||_ dS)Nr rF)r$r%TrA)r'r(rDrPrErFrGrQrRr[conv3rSrOr rTr+r-r.r(s (" zBottleneck.__init__rUcCsL|}||}||}||}|jdur||}||7}||}|Sr0)rQrRr_rOrSrVr-r-r.rXs      zBottleneck.forwardrYrZr-r-r+r.r]|s$ r]c"eZdZdZdfdd ZZS) BasicStemz$The default conv-batchnorm-relu stemr"Nc s4ttjdddddddtdtjdd dS) Nr @)r rcrr r r?Fr@TrArCr*r+r-r.r(s  zBasicStem.__init__r"Nr6r7r8__doc__r(r<r-r-r+r.raracr`) R2Plus1dStemzRR(2+1)D stem is different than the default one as it uses separated 3D convolutionr"Nc sZttjdddddddtdtjdd tjdd d d d ddtd tjdd dS)Nr -)rrcrcrd)rr r Fr@TrArbrBrrr)rrrrCrer+r-r.r(s  zR2Plus1dStem.__init__rfrgr-r-r+r.rjrirjcseZdZ  ddeeeefdeeeee e fde e de dejfde d ed d ffd d Zded efddZ ddeeeefdeeee e fde de de d ejf ddZZS)rFblock conv_makerslayersstem. num_classeszero_init_residualr"Ncstt|d|_||_|j||dd|ddd|_|j||dd|ddd|_|j||dd|ddd|_|j||dd |ddd|_ t d |_ t d |j||_|D]N}t|t jrt jj|jd d d |jdurt j|jdq`t|t jrt j|jdt j|jdq`t|t j rt j|jddt j|jdq`|r|D]}t|trt j|jjdqdSdS)a^Generic resnet video generator. Args: block (Type[Union[BasicBlock, Bottleneck]]): resnet building block conv_makers (List[Type[Union[Conv3DSimple, Conv3DNoTemporal, Conv2Plus1D]]]): generator function for each layer layers (List[int]): number of blocks per layer stem (Callable[..., nn.Module]): module specifying the ResNet stem. num_classes (int, optional): Dimension of the final FC layer. Defaults to 400. zero_init_residual (bool, optional): Zero init bottleneck residual BN. Defaults to False. rbrrr1r r irlZfan_outrS)modeZ nonlinearityNg{Gz?)r'r(r rLrq _make_layerlayer1layer2layer3layer4rDZAdaptiveAvgPool3davgpoolZLinearr[fcmodules isinstancerEinitZkaiming_normal_weightr%Z constant_rFZnormal_r]Zbn3)r*rnrorprqrrrsmr+r-r.r(s<         zVideoResNet.__init__rUcCsT||}||}||}||}||}||}|d}||}|SrJ)rqrxryrzr{r|flattenr})r*rUr-r-r.rXs        zVideoResNet.forwardrrNrMblocksr c Csd}|dks|j||jkr+||}ttj|j||jd|ddt||j}g}|||j||||||j|_td|D] } |||j||qDtj|S)NrF)r$r r%) rLr[r2rDrPrErFappendrange) r*rnrNrMrr rOZ ds_striderpir-r-r.rw s   zVideoResNet._make_layer)rmF)r)r6r7r8typerrKr]rrrIr>listr9rrDr\boolr(rrXrPrwr<r-r-r+r.rsB 4rrnrorprq.weightsprogresskwargsr"cKsT|durt|dt|jdt||||fi|}|dur(||j|dd|S)Nrr categoriesT)rZ check_hash)rlenmetarZload_state_dictZget_state_dict)rnrorprqrrrmodelr-r-r. _video_resnet$s  rrHzKhttps://github.com/pytorch/vision/tree/main/references/video_classificationzThe weights reproduce closely the accuracy of the paper. The accuracies are estimated on video-level with parameters `frame_rate=15`, `clips_per_video=5`, and `clip_len=16`.)Zmin_sizerZrecipeZ_docsc @DeZdZedeedddieddddd id d d d ZeZdS)rz7https://download.pytorch.org/models/r3d_18-b3b3357e.pthprrtZ crop_sizeZ resize_sizeiP5 Kinetics-400gO@g-T@zacc@1zacc@5gK7YD@g"_@Z num_paramsZ_metricsZ_ops _file_sizeurlZ transformsrN r6r7r8rrr _COMMON_METAKINETICS400_V1DEFAULTr-r-r-r.rC$ rc @r)rz7https://download.pytorch.org/models/mc3_18-a90a0ba3.pthrrriPurg{GO@gQU@rgClE@gtVF@rrNrr-r-r-r.rWrrc @r)rz`__. Args: weights (:class:`~torchvision.models.video.R3D_18_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.video.R3D_18_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.video.resnet.VideoResNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.video.R3D_18_Weights :members: r^r r r r )rverifyrrKrrarrrr-r-r.r rcKs4t|}tttgtgdgdt||fi|S)aConstruct 18 layer Mixed Convolution network as in .. betastatus:: video module Reference: `A Closer Look at Spatiotemporal Convolutions for Action Recognition `__. Args: weights (:class:`~torchvision.models.video.MC3_18_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.video.MC3_18_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.video.resnet.VideoResNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.video.MC3_18_Weights :members: r r)rrrrKrrIrarr-r-r.rs rcKr)aConstruct 18 layer deep R(2+1)D network as in .. betastatus:: video module Reference: `A Closer Look at Spatiotemporal Convolutions for Action Recognition `__. Args: weights (:class:`~torchvision.models.video.R2Plus1D_18_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.video.R2Plus1D_18_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.video.resnet.VideoResNet`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.video.R2Plus1D_18_Weights :members: r^r)rrrrKr>rjrr-r-r.rrr) _ModelURLs)rrr)6collections.abcr functoolsrtypingrrrrZtorch.nnrDZtorchrZtransforms._presetsr utilsr Z_apir rr_metarZ_utilsrr__all__rErrPr>rIr\rKr]rarjrrrr9rrrrrrrrrrrrZ model_urlsr-r-r-r.sx       %1 ^   *#*#* $