o ϯi@sddlmZmZmZddlZddlmZmZmZm Z m Z m Z ddl m Z dde ee e e e e e e e e d Zdddddd d d d d d d d d Zd d dZGdddZdS))AnyDictOptionalN)_AUGMENTATION_MAX_LEVELAugmentTransform_decreasing_int_to_arg_decreasing_to_arg_increasing_magnitude_to_arg"_increasing_randomly_negate_to_arg) OpSampler) AutoContrastEqualizeRotate PosterizeSolarizeShearXShearY TranslateX TranslateYAdjustSaturationAdjustContrastAdjustBrightnessAdjustSharpness)r)r)r)rg333333?)rgUUUUUU?)皙?g?floatr)sampling_data_type sampling_minc@sreZdZdZ      ddededed ed eeee fd eeee fd dfd dZ de j d e j fddZ dS)AugMixa This implements AugMix for video. AugMix generates several chains of augmentations on the original video, which are then mixed together with each other and with the original video to create an augmented video. The input video tensor should have shape (T, C, H, W). AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (https://arxiv.org/pdf/1912.02781.pdf) ?N magnitudealphawidthdepthtransform_hparassampling_hparasreturncst|ts Jd|dkr|tksJdtd|dks!Jd|dks)Jd|_tjjt|g|_tjj ||_ fd d t t D}|dkr`t||d d _dSt|d d d d_dS)a Args: magnitude (int): Magnitude used for transform function. Default is 3. alpha (float): Parameter for choosing mixing weights from the beta and Dirichlet distributions. Default is 1.0. width (int): The number of transformation chains. Default is 3. depth (int): The number of transformations in each chain. If depth is -1, each chain will have a random length between 1 and 3 inclusive. Default is -1. transform_hparas (Optional[Dict[Any]]): Transform hyper parameters. Needs to have key fill. By default, the fill value is (0.5, 0.5, 0.5). 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 SAMPLING_AUGMIX_DEFAULT_HPARAS. zmagnitude must be an intrz magnitude must be between 1 and z inclusivegzalpha must be greater than 0rzwidth must be greater than 0c s*g|]}t|jdttdptdqS)r"uniform)transform_namer$prob level_to_argtransform_max_parasr( sampling_typer))r _magnitude_AUGMIX_LEVEL_TO_ARG_TRANSFORM_AUGMIX_MAX_PARAMSSAMPLING_AUGMIX_DEFAULT_HPARAS).0r,r)selfr(R/home/ubuntu/.local/lib/python3.10/site-packages/pytorchvideo/transforms/augmix.py is z#AugMix.__init__..T) num_sample_op replacementr!)r;randomly_sample_depthr<N) isinstanceintrr1torch distributions dirichlet DirichlettensorbetaBetalistr3keysr augmix_fn)r7r$r%r&r'r(r)transforms_listr8r6r9__init__Cs4     zAugMix.__init__videocCs|j}|j}tj|jtjd}|D] }||||7}q|j tj kr7||d|| tj S||d||S)z Perform AugMix to the input video tensor. Args: video (torch.Tensor): Input video tensor with shape (T, C, H, W). )dtyper) rBsamplerEitemr@zerosshapefloat32rIrMuint8type)r7rLmixing_weightsmmixedmwr8r8r9__call__s  zAugMix.__call__)r!r"r!r#NN)__name__ __module__ __qualname____doc__r?rrrstrrrKr@TensorrYr8r8r8r9r 8s0  Ar )typingrrrr@%pytorchvideo.transforms.augmentationsrrrrr r "pytorchvideo.transforms.transformsr r2r3r4r r8r8r8r9sD