U hA @sddlmZddlmZmZmZmZmZmZm Z m Z ddl m Z ddl mZddlmZddlmZdd lmZmZmZdd lmZdd lmZmZd d dddddgZGddde jZGddde jZ Gddde jZ!Gddde j"Z#Gddde j"Z$Gddde jZ%Gdd d e jZ&Gd!d d e j"Z'e e e#e$fee e ee!e fee(ed"e j"feee)ee'd#d$d%Z*d&ed'd(d)Z+Gd*d d eZ,Gd+ddeZ-Gd,ddeZ.eed-e,j/fd.dd/d0ee,e)ee'd1d2dZ0eed-e-j/fd.dd/d0ee-e)ee'd1d3dZ1eed-e.j/fd.dd/d0ee.e)ee'd1d4dZ2dd5lm3Z3e3e,j/j4e-j/j4e.j/j4d6Z5dS)7)partial)AnyCallableListOptionalSequenceTupleTypeUnionN)Tensor)VideoClassification)_log_api_usage_once)register_modelWeights WeightsEnum)_KINETICS400_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface VideoResNetR3D_18_WeightsMC3_18_WeightsR2Plus1D_18_Weightsr3d_18mc3_18 r2plus1d_18csPeZdZd eeeeeeddfdd ZeeeeeefdddZZ S) Conv3DSimpleN in_planes out_planes midplanesstridepaddingreturncstj||d||dddS)N)r r r FZ in_channelsZ out_channels kernel_sizer#r$biassuper__init__selfr r!r"r#r$ __class__Q/var/www/html/venv/lib/python3.8/site-packages/torchvision/models/video/resnet.pyr+szConv3DSimple.__init__r#r%cCs |||fSNr0r#r0r0r1get_downsample_stride'sz"Conv3DSimple.get_downsample_stride)Nrr __name__ __module__ __qualname__intrr+ staticmethodrr5 __classcell__r0r0r.r1rs rcsLeZdZd eeeeeddfdd ZeeeeeefdddZZS) Conv2Plus1DrNrc s`ttj||dd||fd||fddt|tjddtj||d|ddf|ddfdddS) Nrr r rrFr'r#r$r(TZinplacer rrr*r+nnConv3d BatchNorm3dReLUr,r.r0r1r+-s( zConv2Plus1D.__init__r2cCs |||fSr3r0r4r0r0r1r5>sz!Conv2Plus1D.get_downsample_stride)rr) r7r8r9r:r+r;rr5r<r0r0r.r1r=,sr=csPeZdZd eeeeeeddfdd ZeeeeeefdddZZ S) Conv3DNoTemporalNrrcs(tj||dd||fd||fdddS)Nr>rrFr&r)r,r.r0r1r+DszConv3DNoTemporal.__init__r2cCs d||fSNrr0r4r0r0r1r5Qsz&Conv3DNoTemporal.get_downsample_stride)Nrrr6r0r0r.r1rGCs rGcsReZdZdZd eeedejfeeejddfdd Z e e ddd Z Z S) BasicBlockrN.inplanesplanes conv_builderr# downsampler%cs||ddd|ddd|}tt|||||t|tjdd|_t||||t||_tjdd|_||_ ||_ dS)Nr Tr@) r*r+rC SequentialrErFconv1conv2relurNr#r-rKrLrMr#rNr"r.r0r1r+Zs(   zBasicBlock.__init__xr%cCsB|}||}||}|jdk r,||}||7}||}|Sr3)rPrQrNrRr-rUZresidualoutr0r0r1forwardms     zBasicBlock.forward)rN r7r8r9 expansionr:rrCModulerr+r rXr<r0r0r.r1rIVs rIcsReZdZdZd eeedejfeeejddfdd Z e e dd d Z Z S) BottleneckrN.rJcst||ddd|ddd|}ttj||dddt|tjdd|_t|||||t|tjdd|_ttj|||j dddt||j |_ tjdd|_ ||_ ||_ dS)Nr rF)r'r(Tr@)r*r+rCrOrDrErFrPrQrZconv3rRrNr#rSr.r0r1r+~s& (   zBottleneck.__init__rTcCsL|}||}||}||}|jdk r6||}||7}||}|Sr3)rPrQr^rNrRrVr0r0r1rXs      zBottleneck.forward)rNrYr0r0r.r1r\{s r\cs&eZdZdZddfdd ZZS) BasicStemz$The default conv-batchnorm-relu stemNr%c s4ttjdddddddtdtjdd dS) Nr @)r rbrrrr>Fr?Tr@rBr-r.r0r1r+s  zBasicStem.__init__r7r8r9__doc__r+r<r0r0r.r1r_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 -)rrbrbrc)rr r Fr?Tr@rarArrr)rrrrBrdr.r0r1r+s  zR2Plus1dStem.__init__rer0r0r.r1rgsrgc 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 |jdk rbt 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. rarrr4rr iriZfan_outrR)modeZ nonlinearityNg{Gz?)r*r+rrKrn _make_layerlayer1layer2layer3layer4rCZAdaptiveAvgPool3davgpoolZLinearrZfcmodules isinstancerDinitZkaiming_normal_weightr(Z constant_rEZnormal_r\Zbn3)r-rkrlrmrnrorpmr.r0r1r+s2       zVideoResNet.__init__rTcCsT||}||}||}||}||}||}|d}||}|SrH)rnrurvrwrxryflattenrz)r-rUr0r0r1rXs        zVideoResNet.forwardr)rkrMrLblocksr#r%c Csd}|dks|j||jkrV||}ttj|j||jd|ddt||j}g}|||j||||||j|_td|D]} |||j||qtj|S)NrF)r'r#r() rKrZr5rCrOrDrEappendrange) r-rkrMrLrr#rNZ ds_stridermir0r0r1rt s  zVideoResNet._make_layer)rjF)r)r7r8r9r r rIr\rrrGr=rr:rrCr[boolr+r rXrOrtr<r0r0r.r1rs* 4.)rkrlrmrnweightsprogresskwargsr%cKsP|dk rt|dt|jdt||||f|}|dk rL||j|dd|S)Nro categoriesT)rZ check_hash)rlenmetarZload_state_dictZget_state_dict)rkrlrmrnrrrmodelr0r0r1 _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 @s@eZdZedeedddeddddd id d d d ZeZdS)rz7https://download.pytorch.org/models/r3d_18-b3b3357e.pthprrqZ crop_sizeZ resize_sizeiP5 Kinetics-400gO@g-T@zacc@1zacc@5gK7YD@g"_@Z num_paramsZ_metricsZ_ops _file_sizeurlZ transformsrN r7r8r9rrr _COMMON_METAKINETICS400_V1DEFAULTr0r0r0r1rBs c @s@eZdZedeedddeddddd id d d d ZeZdS)rz7https://download.pytorch.org/models/mc3_18-a90a0ba3.pthrrriPurg{GO@gQU@rgClE@gtVF@rrNrr0r0r0r1rVs c @s@eZdZedeedddeddddd 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)rverifyrrIrr_rrrr0r0r1r~s  cKs4t|}tttgtgdddddgt||f|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)rrrrIrrGr_rr0r0r1rs  cKs.t|}tttgdddddgt||f|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)rrrrIr=rgrr0r0r1rs  ) _ModelURLs)rrr)6 functoolsrtypingrrrrrrr r Ztorch.nnrCZtorchr Ztransforms._presetsr utilsrZ_apirrr_metarZ_utilsrr__all__rDrrOr=rGr[rIr\r_rgrr:rrrrrrrrrrrrZ model_urlsr0r0r0r1sr (      %1 _   $#$#$$