o ϯiD@sdZddlmZmZmZmZmZddlZddlZddl m m Z ddl mZdZddZdejd ejfd d Zdejd ejfd d Zdejd ejfddZdejded ejfddZdejded ejfddZdejded ejfddZdejded ejfddZdejded ejfddZdejded ejfddZdejdefddZdejdefdd Zdejdefd!d"Zdejdefd#d$Zdejdefd%d&Z d'ed efd(d)Z!d*e"d+eeefd efd,d-Z#d*e"d+eeefd eefd.d/Z$d*e"d+ee"e"fd ee"fd0d1Z%d*e"d+eeefd eefd2d3Z&eeeeeeeeeeeeee d4Z'e$e$e$e$ddde$e%e&e$e$e$e$d4Z(d5d5d5d5dddd6d7d8d9d9d:d:d4Z)d;diZ+Gd?d@d@Z,dS)AzAVideo transforms that are used for advanced augmentation methods.)AnyCallableDictOptionalTupleN)InterpolationMode cCsd|vsJddS)z0 Check if kwargs contains key ``fill``. fillzNeed to have fill in kwargs.N)kwargsr r Y/home/ubuntu/.local/lib/python3.10/site-packages/pytorchvideo/transforms/augmentations.py_check_fill_argsr videoreturncKtjj|S)z Maximize contrast of a video by remapping its pixels per channel so that the lowest becomes black and the lightest becomes white. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). ) torchvision transforms functional autocontrastrr r r r _autocontrastsrcKsH|jtjkr|j}|dtj}tjj|d|Stjj|S)z Equalize the histogram of a video by applying a non-linear mapping to the input in order to create a uniform distribution of grayscale values in the output. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). )dtypetorchuint8torrrequalize)rr video_typer r r _equalize#s rcKr)zv Invert the colors of a video. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). )rrrinvertrr r r _invert2sr factorcKs$t|tjjj|||dtjdS)z Rotate the image by angle. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): The rotation angle value in degrees, counter-clockwise. r r interpolation)r rrrrotaterBILINEARrr!r r r r _rotate<sr'cKs4|jtjkrtjj|t|dStjj||S)z Solarize an video by inverting all pixel values above a threshold. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). go@)rrrrrrsolarizeintr&r r r _solarizeJs r*cKtjj||S)aP Adjust contrast of an a video. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): How much to adjust the contrast. Can be any non-negative number. 0 gives a solid gray video, 1 gives the original video while 2 increases the contrast by a factor of 2. )rrradjust_contrastr&r r r _adjust_contrastW r-cKr+)aG Adjust the saturation of a video. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): How much to adjust the saturation. 0 will give a black and white video, 1 will give the original video while 2 will enhance the saturation by a factor of 2. )rrradjust_saturationr&r r r _adjust_saturationdr.r0cKr+)aX Adjust brightness of a video. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). sharpness_factor (float): How much to adjust the sharpness. Can be any non-negative number. 0 gives a blurred video, 1 gives the original video while 2 increases the sharpness by a factor of 2. )rrradjust_brightnessr&r r r _adjust_brightnessqr.r2cKr+)aQ Adjust the sharpness of a video. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): How much to adjust the sharpness. Can be any non-negative number. 0 gives a blurred video, 1 gives the original video while 2 increases the sharpness by a factor of 2. )rrradjust_sharpnessr&r r r _adjust_sharpness~r.r4cKsX|dkr|S|jtjkr$|j}|dtj}tjj||d|Stjj||S)z Posterize an image by reducing the number of bits for each color channel. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): The number of bits to keep for each channel (0-8). r)rrrrrrr posterize)rr!r rr r r _posterizes r7cKs<t||d|d}tj|d||dddg|dddS)z Shear the video along the horizontal axis. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): How much to shear along the horizontal axis using the affine matrix. rr bilinearr"r sizeF_taffinerr!r translation_offsetr r r _shear_x rBcKs<t||d|d}tj|ddd|d|g|dddS)z Shear the video along the vertical axis. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): How much to shear along the vertical axis using the affine matrix. r9r:rr r;r"r<r@r r r _shear_yrCrEcKs8t|||d}tj|dd|dddg|dddS)z Translate the video along the vertical axis. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): How much (relative to the image size) to translate along the vertical axis. rDr:rr r;r"r<r@r r r _translate_x rFcKs8t|||d}tj|ddddd|g|dddS)z Translate the video along the vertical axis. Args: video (torch.Tensor): Video tensor with shape (T, C, H, W). factor (float): How much (relative to the image size) to translate along the horizontal axis. r8r:rr r;r"r<r@r r r _translate_yrGrH magnitudecCstddkr |S| S)z` Negate input value with 50% chance. Args: magnitude (float): Input value. r:?)rranditem)rIr r r _randomly_negatesrMlevelparamscCs|t|d}|d|fS)ap Convert level to transform magnitude. This assumes transform magnitude increases linearly with level. Args: level (int): Level value. params (Tuple[float, float]): Params contains two values: 1) Base transform magnitude when level is 0; 2) Maxmimum increasing in transform magnitude when level is at Maxmimum. r:r_AUGMENTATION_MAX_LEVELrNrOrIr r r _increasing_magnitude_to_arg rScCs"|t|d}|dt|fS)a Convert level to transform magnitude. This assumes transform magnitude increases (or decreases with 50% chance) linearly with level. Args: level (int): Level value. params (Tuple[float, float]): Params contains two values: 1) Base transform magnitude when level is 0; 2) Maxmimum increasing in transform magnitude when level is at maxmimum. r:r)rQrMrRr r r "_increasing_randomly_negate_to_args rUcCs"|t|d}|dt|fS)a Convert level to transform magnitude. This assumes transform magnitude decreases linearly with level. The return value is converted to int. Args: level (int): Level value. params (Tuple[float, float]): Params contains two values: 1) Base transform magnitude when level is 0; 2) Maxmimum decreasing in transform magnitude when level is at maxmimum. r:r)rQr)rRr r r _decreasing_int_to_args rVcCs|t|d}|d|fS)ap Convert level to transform magnitude. This assumes transform magnitude decreases linearly with level. Args: level (int): Level value. params (Tuple[float, float]): Params contains two values: 1) Base transform magnitude when level is 0; 2) Maxmimum decreasing in transform magnitude when level is at maxmimum. r:rrPrRr r r _decreasing_to_arg#rTrW)AdjustBrightnessAdjustContrastAdjustSaturationAdjustSharpness AutoContrastEqualizeInvertRotate PosterizeSolarizeShearXShearY TranslateX TranslateY)r:g?)r)rg)r:r:)rg333333?)rg? sampling_stdrJr )rJrJrJc@seZdZ        ddedededeeeefd eeeefd eeee fd eeee fd ed eeee fddfddZ defddZ de jde jfddZdS)AugmentTransformrrJNgaussiantransform_namerIprobname_to_transform_func level_to_argtransform_max_parastransform_hparas sampling_typesampling_hparasrc Cs |dvsJ|p t}|p t}|pt}|pt|_||_| pt|_d|jvs&J|jdkr2d|jvs2J|jdkrjd|jvs>Jd|jvsEJ|jddkrWt|jdt sVJn|jdd krjt|jdt t fsjJ||vspJt |_ ||_ ||_|||_|||_|||_||_||_d S) aM The AugmentTransform composes a video transform that performs augmentation based on a maximum magnitude. AugmentTransform also offers flexible ways to generate augmentation magnitude based on different sampling strategies. Args: transform_name (str): The name of the video transform function. magnitude (int): Magnitude used for transform function. prob (float): The probablity of applying each transform function. name_to_transform_func (Optional[Dict[str, Callable]]): A Dictionary that contains mapping of the transform name to the transform function. level_to_arg (Optional[Dict[str, Callable]]): A Dictionary that contains mapping of the transform name to its level function, which converts the the magnitude to the transform function arguments. transform_max_paras (Optional[Dict[str, Tuple]]): A Dictionary that contains mapping of the transform name to its maximum transform magnitude. transform_hparas (Optional[Dict[Any]]): Transform hyper parameters. Needs to have key fill. By default, it uses transform_default_hparas. sampling_type (str): Sampling method for magnitude of transform. It should be either gaussian or uniform. sampling_hparas (Optional[Dict[Any]]): Hyper parameters for sampling. If gaussian sampling is used, it needs to have key sampling_std. By default, it uses transform_default_hparas. )rjuniformr rjrhrssampling_data_type sampling_minr)floatN)_NAME_TO_TRANSFORM_FUNC _LEVEL_TO_ARG_TRANSFORM_MAX_PARAMSTRANSFORM_DEFAULT_HPARASrprqSAMPLING_DEFAULT_HPARASrr isinstancer)rvrQ max_levelrkrI transform_fnlevel_fn level_parasrl) selfrkrIrlrmrnrorprqrrr r r __init__ss4 &         zAugmentTransform.__init__c Cs|jdkrtdt|jtj|j|jdddS|jdkrV|jddkr6tj |jd |jd ddS|jdd krRtj dd|j|jd |jd St d t ) z7 Get magnitude based on sampling type. rjrrh)r:)r=rsrtr)rur:rvz2sampling_data_type must be either 'int' or 'float') rqmaxminr}rnormalrIrrrLrandintrK ValueErrorNotImplementedError)rr r r _get_magnitudes6  zAugmentTransform._get_magnitudercCsVtd|jkr |S|}|jdur|||jnd}|j|g|Ri|jS)z The input is a video tensor. Args: video (torch.Tensor): Input video tensor with shape (T, C, H, W). r:Nr ) rrKrLrlrrrr~rp)rrrI level_argsr r r __call__s zAugmentTransform.__call__)rrJNNNNrjN)__name__ __module__ __qualname__strr)rvrrrrrrrrTensorrr r r r rirs@   Bri)-__doc__typingrrrrrrr(torchvision.transforms.functional_tensorrfunctional_tensorr>!torchvision.transforms.functionalrrQr rrrr rvr'r*r-r0r2r4r7rBrErFrHrMr)rSrUrVrWrwrxryr{rzrir r r r s           ""