o ϯi3@srddlmZmZmZmZmZddlZddlZddl Z GdddZ Gdddej j ZGdddej j ZGd d d ej j ZGd d d ej j ZGd ddej j ZGdddej j ZGddde jjZGdddej j ZGdddej j ZGdddej j ZGdddej j ZGdddej j ZGdddej j ZGdd d ej j ZGd!d"d"ej j ZdS)#)CallableDictListOptionalTupleNc@sHeZdZdZdedefddZdeeej fdeeej ffdd Z d S) ApplyTransformToKeyat Applies transform to key of dictionary input. Args: key (str): the dictionary key the transform is applied to transform (callable): the transform that is applied Example: >>> transforms.ApplyTransformToKey( >>> key='video', >>> transform=UniformTemporalSubsample(num_video_samples), >>> ) key transformcCs||_||_dSN)_key _transform)selfrr rV/home/ubuntu/.local/lib/python3.10/site-packages/pytorchvideo/transforms/transforms.py__init__s zApplyTransformToKey.__init__xreturncCs|||j||j<|Sr )r r r rrrr__call__szApplyTransformToKey.__call__N) __name__ __module__ __qualname____doc__strrrrtorchTensorrrrrrr s*rcsLeZdZdZdeffdd ZdeeejfdeeejffddZ Z S) RemoveKeyz} Removes the given key from the input dict. Useful for removing modalities from a video clip that aren't needed. rcst||_dSr )superrr )r r __class__rrr(s  zRemoveKey.__init__rrcCs|j|vr ||j=|SzQ Args: x (Dict[str, torch.Tensor]): video clip dict. )r rrrrr,s zRemoveKey.__call__) rrrrrrrrrr __classcell__rrrrr"s.rcsBeZdZdZd dedeffdd Zdejdejfd d ZZ S) UniformTemporalSubsamplezf ``nn.Module`` wrapper for ``pytorchvideo.transforms.functional.uniform_temporal_subsample``. num_samples temporal_dimcst||_||_dS)z Args: num_samples (int): The number of equispaced samples to be selected temporal_dim (int): dimension of temporal to perform temporal subsample. N)rr _num_samples _temporal_dim)r r$r%rrrr;s  z!UniformTemporalSubsample.__init__rrcCtjj||j|jS[ Args: x (torch.Tensor): video tensor with shape (C, T, H, W). ) pytorchvideo transforms functionaluniform_temporal_subsampler&r'rrrrforwardE z UniformTemporalSubsample.forwardr#) rrrrintrrrr/r!rrrrr"6s r"cs@eZdZdZd deedeffdd Zdejfdd Z Z S) UniformTemporalSubsampleRepeatedzs ``nn.Module`` wrapper for ``pytorchvideo.transforms.functional.uniform_temporal_subsample_repeated``. r# frame_ratiosr%cst||_||_dSr )rr _frame_ratiosr')r r4r%rrrrUs  z)UniformTemporalSubsampleRepeated.__init__rcCr(r))r+r,r-#uniform_temporal_subsample_repeatedr5r'rrrrr/Zr0z(UniformTemporalSubsampleRepeated.forwardr1 rrrrrr2rrrr/r!rrrrr3Osr3csHeZdZdZ d dededeffdd Zd ejd ejfd d Z Z S)ShortSideScalez\ ``nn.Module`` wrapper for ``pytorchvideo.transforms.functional.short_side_scale``. bilinearpytorchsize interpolationbackendc t||_||_||_dSr )rr_size_interpolation_backend)r r;r<r=rrrri  zShortSideScale.__init__rrcCstjj||j|j|jSr))r+r,r-short_side_scaler?r@rArrrrr/qszShortSideScale.forwardr9r: rrrrr2rrrrr/r!rrrrr8dsr8c sNeZdZdZ  ddedededeffdd Zd ejd ejfd d Z Z S)RandomShortSideScalez ``nn.Module`` wrapper for ``pytorchvideo.transforms.functional.short_side_scale``. The size parameter is chosen randomly in [min_size, max_size]. r9r:min_sizemax_sizer<r=cs&t||_||_||_||_dSr )rr _min_size _max_sizer@rA)r rGrHr<r=rrrrs  zRandomShortSideScale.__init__rrcCs2t|j|jdd}tjj|||j |j S)r*rK) rrandintrIrJitemr+r,r-rCr@rA)r rr;rrrr/s zRandomShortSideScale.forwardrDrErrrrrF{s  rFcsXeZdZdZ d dededeffdd Zd eeej fd eeej ffd d Z Z S)UniformCropVideozX ``nn.Module`` wrapper for ``pytorchvideo.transforms.functional.uniform_crop``. video aug_indexr; video_key aug_index_keycr>r )rrr? _video_key_aug_index_key)r r;rRrSrrrrrBzUniformCropVideo.__init__rrcCs*tjj||j|j||j||j<|Sr )r+r,r- uniform_croprTr?rUrrrrrs zUniformCropVideo.__call__)rPrQ) rrrrr2rrrrrrr!rrrrrOs.rOcs.eZdZdZdejdejffdd ZZS) Normalizea Normalize the (CTHW) video clip by mean subtraction and division by standard deviation Args: mean (3-tuple): pixel RGB mean std (3-tuple): pixel RGB standard deviation inplace (boolean): whether do in-place normalization rrcs0|dddd}t|}|dddd}|S)r*rKr)permuterr/)r rvidrrrr/s zNormalize.forward)rrrrrrr/r!rrrrrWs" rWc6eZdZdZfddZdejdejfddZZS)ConvertFloatToUint8z= Converts a video from dtype float32 to dtype uint8. cttjtj|_dSr )rr torchvisionr,ConvertImageDtyperuint8 convert_funcr rrrr zConvertFloatToUint8.__init__rrcCs*|jtjks|jtjksJd||Sr*z!image must have dtype torch.uint8)dtyperfloathalfrbrrrrr/s zConvertFloatToUint8.forward rrrrrrrr/r!rrrrr] r]cr\)ConvertUint8ToFloatz= Converts a video from dtype uint8 to dtype float32. cr^r )rrr_r,r`rfloat32rbrcrrrrrdzConvertUint8ToFloat.__init__rrcCs|jtjks Jd||Sre)rfrrarbrrrrr/s zConvertUint8ToFloat.forwardrirrrrrkrjrkc>eZdZdZfddZejjdejdejfddZ Z S)MoveChannelRearz= A Scriptable version to perform C X Y Z -> X Y Z C. ctdSr rrrcrrrrzMoveChannelRear.__init__rrcC|gd}|S)g Args: x (torch.Tensor): video tensor whose dimensions are to be permuted. )rKrXrYrrZrrrrr/zMoveChannelRear.forward rrrrrrjit script_methodrr/r!rrrrrn   rncrm)MoveChannelFrontz= A Scriptable version to perform X Y Z C -> C X Y Z. cror rprcrrrrrqzMoveChannelFront.__init__rrcCrr)rs)rYrrKrXrtrrrrr/ruzMoveChannelFront.forwardrvrrrrrzryrzcsveZdZdZ    ddededeeefd eeefd ed ed ed eddffdd Z de j de j fddZ Z S)RandomResizedCropz_ ``nn.Module`` wrapper for ``pytorchvideo.transforms.functional.random_resized_crop``. FTr9 target_height target_widthscale aspect_ratioshiftlog_uniform_ratior< num_triesrNc s>t||_||_||_||_||_||_||_||_ dSr ) rr_target_height _target_width_scale _aspect_ratio_shift_log_uniform_ratior@ _num_tries) r r}r~rrrrr<rrrrrs  zRandomResizedCrop.__init__rc Cs.tjj||j|j|j|j|j|j |j |j S)za Args: x (torch.Tensor): Input video tensor with shape (C, T, H, W). ) r+r,r-random_resized_croprrrrrrr@rrrrrr+szRandomResizedCrop.__call__)FTr9r|)rrrrr2rrgboolrrrrrr!rrrrr{s4     r{cs@eZdZdZdeeffdd ZdejdejfddZ Z S) Permutez- Permutes the dimensions of a video. dimscs6fddttDsJdt|_dS)zZ Args: dims (Tuple[int]): The desired ordering of dimensions. c3s|]}|vVqdSr r).0drrr Gs z#Permute.__init__..z0dims must contain every dimension (0, 1, 2, ...)N)rangelenrr_dims)r rrrrrBs    zPermute.__init__rrcCs |j|jS)rs)rZrrrrrr/Ns zPermute.forwardr7rrrrr=s rc sbeZdZdZ    ddeedeeedede d e f fd d Z d e j d e j fddZ ZS) OpSamplerz Given a list of transforms with weights, OpSampler applies weighted sampling to select n transforms, which are then applied sequentially to the input. NrKFtransforms_listtransforms_prob num_sample_oprandomly_sample_depth replacementcstt|dksJd|dksJd|t|ks!Jd|dur;t|t|ks1Jdt|dks;Jd||_t|durF|ndgt||_||_||_ ||_ dS) a5 Args: transforms_list (List[Callable]): A list of tuples of all available transforms to sample from. transforms_prob (Optional[List[float]]): The probabilities associated with each transform in transforms_list. If not provided, the sampler assumes a uniform distribution over all transforms. They do not need to sum up to one but weights need to be positive. num_sample_op (int): Number of transforms to sample and apply to input. randomly_sample_depth (bool): If randomly_sample_depth is True, then uniformly sample the number of transforms to apply, between 1 and num_sample_op. replacement (bool): If replacement is True, transforms are drawn with replacement. rz)Argument transforms_list cannot be empty.z&Need to sample at least one transform.zMArgument num_sample_op cannot be greater than number of available transforms.NzJArgument transforms_prob needs to have the same length as transforms_list.z4Argument transforms_prob needs to be greater than 0.rK) rrrminrr FloatTensorrrrr)r rrrrrrrrr\s4   zOpSampler.__init__rrcCsT|jrtd|jddn|j}tj|j||jd}|D] }|j||}q|S)zC Args: x (torch.Tensor): Input tensor. rKrL)r) rrrMrrN multinomialrrr)r rdepth index_listindexrrrr/s zOpSampler.forward)NrKFF)rrrrrrrrgr2rrrrr/r!rrrrrVs$ /rc@s&eZdZdZdejdejfddZdS)Div255zS ``nn.Module`` wrapper for ``pytorchvideo.transforms.functional.div_255``. rrcCstjtjjj|S)z Scale clip frames from [0, 255] to [0, 1]. Args: x (Tensor): A tensor of the clip's RGB frames with shape: (C, T, H, W). Returns: x (Tensor): Scaled tensor by dividing 255. )r_r,Lambdar+r-div_255rrrrr/s  zDiv255.forwardN)rrrrrrr/rrrrrsr)typingrrrrr"pytorchvideo.transforms.functionalr+rtorchvision.transformsr_rnnModulerr"r3r8rFrOr,rWr]rkrnrzr{rrrrrrrs(-I