o V h @sddlmZddlmZmZmZddlZddlmZddlm Z ddl m Z ddl m Z d d lmZmZmZd d lmZd d lmZmZgd ZGdddejZGdddejZGdddejZGdddeZeedejfdddddeedededefddZ dS) )partial)AnyCallableOptionalN)nn)Conv3dNormActivation)VideoClassification)_log_api_usage_once)register_modelWeights WeightsEnum)_KINETICS400_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface)S3D S3D_Weightss3dcs@eZdZdededededededejff fdd ZZS) TemporalSeparableConv in_planes out_planes kernel_sizestridepadding norm_layer.c sZtt||d||fd||fd||fd|dt|||ddf|ddf|ddfd|ddS)NrF)rrrbiasr)super__init__r)selfrrrrrr __class__P/var/www/vscode/kcb/lib/python3.10/site-packages/torchvision/models/video/s3d.pyrs(  zTemporalSeparableConv.__init__) __name__ __module__ __qualname__intrrModuler __classcell__r#r#r!r$rs rcsPeZdZdedededededededed ejfffd d Zd d ZZ S)SepInceptionBlock3Drb0_outb1_midb1_outb2_midb2_outb3_outr.c stt||dd|d|_tt||dd|dt||ddd|d|_tt||dd|dt||ddd|d|_ttj ddddt||dd|d|_ dS)Nrrrrr)rrrrrrrrrr) rrrbranch0r Sequentialrbranch1branch2 MaxPool3dbranch3) r rr,r-r.r/r0r1rr!r#r$r7s  zSepInceptionBlock3D.__init__cCs@||}||}||}||}t||||fd}|S)Nr)r5r7r8r:torchcat)r xx0x1x2x3outr#r#r$forwardRs    zSepInceptionBlock3D.forward) r%r&r'r(rrr)rrCr*r#r#r!r$r+6s& r+c sReZdZdZ   ddededeedej j fd dffd d Z d d Z Z S)raWS3D main class. Args: num_class (int): number of classes for the classification task. dropout (float): dropout probability. norm_layer (Optional[Callable]): Module specifying the normalization layer to use. Inputs: x (Tensor): batch of videos with dimensions (batch, channel, time, height, width) 皙?N num_classesdropoutr.returncstt||durttjddd}ttddddd|tjddd d t ddd d |d tdd dd d |tjddd d t d dddddd|t dddd ddd|tjdddd t dd ddddd|t ddddddd|t ddddddd|t ddddddd|t d ddd!ddd|tjddd"d t d#ddd!ddd|t d#d$d d$ddd||_ tj d%d d&|_ ttj|d'tjd(|d d d)d*|_dS)+NgMbP?)epsmomentumr@r )rrr)rr r )rrrr4rr2` r3)r r r )rrri0ipi ii@)rrri@i)r rLrL)rr)piT)rrr)rrr rr BatchNorm3dr6rr9rr+features AvgPool3davgpoolDropoutConv3d classifier)r rFrGrr!r#r$rhsB   z S3D.__init__cCs0||}||}||}tj|dd}|S)N)r r)dim)r\r^rar;mean)r r=r#r#r$rCs   z S3D.forward)rDrEN)r%r&r'__doc__r(floatrrr;rr)rrCr*r#r#r!r$r\s *rc @sFeZdZedeedddddedddd d d d id dd dZeZdS)rz4https://download.pytorch.org/models/s3d-d76dad2f.pth)rWrW)rRrR) crop_size resize_sizezOhttps://github.com/pytorch/vision/tree/main/references/video_classification#s3dzThe weights aim to approximate the accuracy of the paper. The accuracies are estimated on clip-level with parameters `frame_rate=15`, `clips_per_video=1`, and `clip_len=128`.i0~z Kinetics-400gd;OQ@g33333V@)zacc@1zacc@5gv1@gF?@) min_sizemin_temporal_size categoriesrecipe_docs num_params_metrics_ops _file_size)url transformsmetaN) r%r&r'r rr rKINETICS400_V1DEFAULTr#r#r#r$rs0r pretrained)weightsT)ryprogressryrzkwargsrHcKsVt|}|durt|dt|jdtdi|}|dur)||j|dd|S)aConstruct Separable 3D CNN model. Reference: `Rethinking Spatiotemporal Feature Learning `__. .. betastatus:: video module Args: weights (:class:`~torchvision.models.video.S3D_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.video.S3D_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.S3D`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.video.S3D_Weights :members: NrFrlT)rz check_hashr#)rverifyrlenrurload_state_dictget_state_dict)ryrzr{modelr#r#r$rs r)! functoolsrtypingrrrr;rtorchvision.ops.miscrtransforms._presetsr utilsr _apir r r_metar_utilsrr__all__r6rr)r+rrrvboolrr#r#r#r$s$       &>.