a 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$Gddde j Z%e&ee!e"fee&eeeefe'e(ede j feee)ee%dd d!Z*d"ed#d$d%Z+Gd&d'd'eZ,Gd(d)d)eZ-Gd*d+d+eZ.eed,e,j/fd-dd.d/ee,e)ee%d0d1d2Z0eed,e-j/fd-dd.d/ee-e)ee%d0d3d4Z1eed,e.j/fd-dd.d/ee.e)ee%d0d5d6Z2d d7lm3Z3e3e,j/j4e-j/j4e.j/j4d8Z5dS)9)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_18csPeZdZd eeeeeeddfdd ZeeeeeefdddZZ 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__M/var/www/auris/lib/python3.9/site-packages/torchvision/models/video/resnet.pyr(szConv3DSimple.__init__r r"cCs |||fSNr-r r-r-r.get_downsample_stride(sz"Conv3DSimple.get_downsample_stride)Nrr __name__ __module__ __qualname__intrr( staticmethodtupler2 __classcell__r-r-r+r.rs rcsLeZdZd eeeeeddfdd ZeeeeeefdddZZS) Conv2Plus1DrNrc s`ttj||dd||fd||fddt|tjddtj||d|ddf|ddfdddS) Nrr r rrFr$r r!r%TZinplacer rrr'r(nnConv3d BatchNorm3dReLUr)r+r-r.r(.s zConv2Plus1D.__init__r/cCs |||fSr0r-r1r-r-r.r2?sz!Conv2Plus1D.get_downsample_stride)rr) r4r5r6r7r(r8r9r2r:r-r-r+r.r;-sr;csPeZdZd eeeeeeddfdd ZeeeeeefdddZZ S) Conv3DNoTemporalNrrcs(tj||dd||fd||fdddS)Nr<rrFr#r&r)r+r-r.r(EszConv3DNoTemporal.__init__r/cCs d||fSNrr-r1r-r-r.r2Rsz&Conv3DNoTemporal.get_downsample_stride)Nrrr3r-r-r+r.rEDs rEcsReZdZdZd eeedejfeeejddfdd Z e e ddd Z Z S) BasicBlockrN.inplanesplanes conv_builderr downsampler"cs||ddd|ddd|}tt|||||t|tjdd|_t||||t||_tjdd|_||_ ||_ dS)Nr Tr>) r'r(rA SequentialrCrDconv1conv2relurLr r*rIrJrKr rLrr+r-r.r([s( zBasicBlock.__init__xr"cCsB|}||}||}|jdur,||}||7}||}|Sr0)rNrOrLrPr*rSZresidualoutr-r-r.forwardns     zBasicBlock.forward)rN r4r5r6 expansionr7rrAModulerr(rrVr:r-r-r+r.rGWs rGcsReZdZdZd eeedejfeeejddfdd Z e e dd d Z Z S) BottleneckrN.rHcst||ddd|ddd|}ttj||dddt|tjdd|_t|||||t|tjdd|_ttj|||j dddt||j |_ tjdd|_ ||_ ||_ dS)Nr rF)r$r%Tr>)r'r(rArMrBrCrDrNrOrXconv3rPrLr rQr+r-r.r(s ("zBottleneck.__init__rRcCsL|}||}||}||}|jdur6||}||7}||}|Sr0)rNrOr\rLrPrTr-r-r.rVs      zBottleneck.forward)rNrWr-r-r+r.rZ|s rZcs&eZdZdZddfdd ZZS) BasicStemz$The default conv-batchnorm-relu stemNr"c s4ttjdddddddtdtjdd dS) Nr @)r r`rr r r<Fr=Tr>r@r*r+r-r.r(s  zBasicStem.__init__r4r5r6__doc__r(r:r-r-r+r.r]sr]cs&eZdZdZddfdd ZZS) R2Plus1dStemzRR(2+1)D stem is different than the default one as it uses separated 3D convolutionNr^c sZttjdddddddtdtjdd tjdd d d d ddtd tjdd dS)Nr -)rr`r`ra)rr r Fr=Tr>r_r?rrr)rrrr@rbr+r-r.r(s  zR2Plus1dStem.__init__rcr-r-r+r.resrec seZdZdeeeefeeeee e fe e e dejfe eddfdd Zeedd d Zdeeeefeeee e fe e e ejd d dZZS)rF.N)block conv_makerslayersstem num_classeszero_init_residualr"cstt|d|_||_|j||dd|ddd|_|j||dd|ddd|_|j||dd|ddd|_|j||dd |ddd|_ t d |_ t d |j||_|D]}t|t jrt jj|jd d d |jdurbt j|jdqt|t jr4t j|jdt j|jdqt|t j rt j|jddt j|jdq|r|D]$}t|trrt j|jjdqrdS)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. r_rrr1r r irgZfan_outrP)modeZ nonlinearityNg{Gz?)r'r(r rIrl _make_layerlayer1layer2layer3layer4rAZAdaptiveAvgPool3davgpoolZLinearrXfcmodules isinstancerBinitZkaiming_normal_Zweightr%Z constant_rCZnormal_rZZbn3)r*rirjrkrlrmrnmr+r-r.r(s2       zVideoResNet.__init__rRcCsT||}||}||}||}||}||}|d}||}|SrF)rlrsrtrurvrwflattenrx)r*rSr-r-r.rVs        zVideoResNet.forwardr)rirKrJblocksr r"c Csd}|dks|j||jkrV||}ttj|j||jd|ddt||j}g}|||j||||||j|_td|D]} |||j||qtj|S)NrF)r$r r%) rIrXr2rArMrBrCappendrange) r*rirKrJr~r rLZ ds_striderkir-r-r.rr s  zVideoResNet._make_layer)rhF)r)r4r5r6typerrGrZrrrEr;listr7rrArYboolr(rrVrMrrr:r-r-r+r.rs* 4r.)rirjrkrlweightsprogresskwargsr"cKsT|durt|dt|jdt||||fi|}|durP||j|dd|S)Nrm categoriesT)rZ check_hash)rlenmetarZload_state_dictZget_state_dict)rirjrkrlrrrmodelr-r-r. _video_resnet$s r)rrzKhttps://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 @sDeZdZedeedddieddddd id d d d ZeZdS)rz7https://download.pytorch.org/models/r3d_18-b3b3357e.pthprroZ crop_sizeZ resize_sizeiP5 Kinetics-400gO@g-T@zacc@1zacc@5gK7YD@g"_@Z num_paramsZ_metricsZ_ops _file_sizeurlZ transformsrN r4r5r6rrr _COMMON_METAKINETICS400_V1DEFAULTr-r-r-r.rCs" rc @sDeZdZedeedddieddddd id d d d ZeZdS)rz7https://download.pytorch.org/models/mc3_18-a90a0ba3.pthrrriPurg{GO@gQU@rgClE@gtVF@rrNrr-r-r-r.rWs" rc @sDeZdZedeedddieddddd id d d d ZeZdS)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 )rverifyrrGrr]rrrr-r-r.rs 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)rrrrGrrEr]rr-r-r.rs rcKs.t|}tttgdgdt||fi|S)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)rrrrGr;rerr-r-r.rs r) _ModelURLs)rrr)6collections.abcr functoolsrtypingrrrrZtorch.nnrAZtorchrZtransforms._presetsr utilsr Z_apir rrZ_metar_utilsrr__all__rBrrMr;rErYrGrZr]rerrrr7rrrrrrrrrrrrZ model_urlsr-r-r-r.sf        %1 _   $#$#$$