o h@n @sddlZddlmZddlmZddlZddlmZddlmZ m Z gdZ ded e d e d e d eee f d dZGdddeZGdddejjZGdddejjZGdddejjZGdddejjZdS)N)Enum)Optional)Tensor) functionalInterpolationMode)AutoAugmentPolicy AutoAugment RandAugmentTrivialAugmentWideAugMiximgop_name magnitude interpolationfillc Cs|dkrtj|dddgdtt|dg||ddgd}|S|dkr>tj|dddgddtt|g||ddgd}|S|dkrVtj|dt|dgd|ddg|d}|S|d krntj|ddt|gd|ddg|d}|S|d kr}tj||||d }|S|d krt|d|}|S|d krt|d|}|S|dkrt |d|}|S|dkrt |d|}|S|dkrt |t|}|S|dkrt ||}|S|dkrt |}|S|dkrt|}|S|dkrt|}|S|dkr |Std|d)NShearXr?)angle translatescaleshearrrcenterShearY TranslateX)rrrrrr TranslateYRotaterr BrightnessColorContrast Sharpness PosterizeSolarize AutoContrastEqualizeInvertIdentityzThe provided operator  is not recognized.)FZaffinemathdegreesatanintrotateZadjust_brightnessZadjust_saturationZadjust_contrastZadjust_sharpnessZ posterizeZsolarizeZ autocontrastZequalizeinvert ValueError)r rrrrr2h/home/www/facesmatcher.com/frenv_anti/lib/python3.10/site-packages/torchvision/transforms/autoaugment.py _apply_op sC6 + !       r4c@seZdZdZdZdZdZdS)rzoAutoAugment policies learned on different datasets. Available policies are IMAGENET, CIFAR10 and SVHN. ZimagenetZcifar10ZsvhnN)__name__ __module__ __qualname____doc__IMAGENETCIFAR10SVHNr2r2r2r3r]s rc seZdZdZejejdfdededee e ddffdd Z dede e e e e eefe e e eefffd d Zd ed e eefdee e eefffd dZedede eeeffddZdedefddZde fddZZS)r a?AutoAugment data augmentation method based on `"AutoAugment: Learning Augmentation Strategies from Data" `_. If the image is torch Tensor, it should be of type torch.uint8, and it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. If img is PIL Image, it is expected to be in mode "L" or "RGB". Args: policy (AutoAugmentPolicy): Desired policy enum defined by :class:`torchvision.transforms.autoaugment.AutoAugmentPolicy`. Default is ``AutoAugmentPolicy.IMAGENET``. interpolation (InterpolationMode): Desired interpolation enum defined by :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. fill (sequence or number, optional): Pixel fill value for the area outside the transformed image. If given a number, the value is used for all bands respectively. Npolicyrrreturncs,t||_||_||_|||_dSN)super__init__r<rr _get_policiespolicies)selfr<rr __class__r2r3r@ys zAutoAugment.__init__cCsF|tjkr gdS|tjkrgdS|tjkrgdStd|d)N)))r#皙?)r333333? )r$rHr%rHNr&皙?Nr&rHN))r#rH)r#rHr&rFN)r$皙?)rTrrOrG))r$rHrP))r#rOrKr&rN))rrUrX)r$rHrG)rP)r#rFrR)rWr rFr))rrFrIrP))r&rNrNr'rHNrY)r rHrV)r!rrG)rW)r r))r rOrG)r$rOrQ))r"rFrQr\))rrHrKrY)rZrPrSrJr[r]rM))r'皙?N)r!rUrR))rffffff?r^)r333333?rI))r"rOr)r"?rX))r?rGrrarI))r%rdNr&rcN))rrUrQ)r#rbrQ))r rFrX)rrHrQ))r"rbrI)rrarI)rP)r&rdN))r!rHrQ)r"rHrK))r rarQ)rrdrG))r&rbN)r%rFN))rrFrX)r"rUrR))rrcrR)r rUrG))r$rdr^)r'rN)r&rUNrL)rgrP))r rcrIrP)r%rON)r$rUrG))rr`rX)r rar))r$rFrKr%rcN))rrcrIre)ri)r$rOrX)rNr_)reri))rrcrV)r'rUN)rrcrGr'raN)rP)r$rHrRr'rcNrPrP)rrcrX)rjrh)rk)r'rFN))rrcrK)r$rUrR)rnrhro)rj)r$rbrX))rrOrGrl)rf)rrHrRrm))r!rbrXrrOrV)r'rON)rrr^))rrarR)r$rFrG)r\rp))rrbrQ)rrcrX))rr`rRr\))r$rar^)rrHrQ))rrOrVrq))rrarI)rrOrX))rrOrK)r%raN))rrar^r_zThe provided policy r))rr9r:r;r1)rCr<r2r2r3rAs   zAutoAugment._get_policiesnum_bins image_sizecCstdd|dftdd|dftdd|d|dftdd|d|dftdd|dftdd|dftdd|dftdd|dftdd|dfd t||dd d ftd d|d ftdd ftdd ftdd fd S)NrrbTt?rr>@rcrGrVFo@)rrrrrrr r!r"r#r$r%r&r')torchlinspacearangeroundr.tensorrCrrrsr2r2r3_augmentation_spaces$   zAutoAugment._augmentation_space transform_numcCs4tt|d}td}tdd}|||fS)zGet parameters for autoaugment transformation Returns: params required by the autoaugment transformation r)r^r^)r.rwrandintitemZrand)r~Z policy_idprobssignsr2r2r3 get_paramss   zAutoAugment.get_paramsr cCs|j}t|\}}}t|tr*t|ttfrt|g|}n |dur*dd|D}|t|j \}}}| d||f} t |j |D]7\} \} } } || | kr{| | \}}| durct||  nd}|rq|| dkrq|d9}t || ||j|d}qD|S) z img (PIL Image or Tensor): Image to be transformed. Returns: PIL Image or Tensor: AutoAugmented image. NcSg|]}t|qSr2float.0fr2r2r3 z'AutoAugment.forward.. rrr)rr*get_dimensions isinstancerr.rrlenrBr} enumeraterr4r)rCr rchannelsheightwidthZ transform_idrrop_metairpZ magnitude_id magnitudessignedrr2r2r3forwards$   zAutoAugment.forwardcCs|jjd|jd|jdS)Nz(policy=, fill=))rEr5r<r)rCr2r2r3__repr__szAutoAugment.__repr__)r5r6r7r8rr9rNEARESTrlistrr@tuplestrr.rAdictrboolr} staticmethodrrr __classcell__r2r2rDr3r hs0  * .Z r c seZdZdZdddejdfdededed ed eee d df fd d Z dede eefd e e e eefffddZded efddZd e fddZZS)r a~RandAugment data augmentation method based on `"RandAugment: Practical automated data augmentation with a reduced search space" `_. If the image is torch Tensor, it should be of type torch.uint8, and it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. If img is PIL Image, it is expected to be in mode "L" or "RGB". Args: num_ops (int): Number of augmentation transformations to apply sequentially. magnitude (int): Magnitude for all the transformations. num_magnitude_bins (int): The number of different magnitude values. interpolation (InterpolationMode): Desired interpolation enum defined by :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. fill (sequence or number, optional): Pixel fill value for the area outside the transformed image. If given a number, the value is used for all bands respectively. r^rINnum_opsrnum_magnitude_binsrrr=cs,t||_||_||_||_||_dSr>)r?r@rrrrr)rCrrrrrrDr2r3r@2s  zRandAugment.__init__rrrscCstddftdd|dftdd|dftdd|d|dftdd|d|dftdd|dftdd |dftdd |dftdd |dftdd |dfd t||dd dftd d|dftddftddfd S)NrFrbTrtrrrurcrGrVrvr(rrrrrrr r!r"r#r$r%r&rwr{rxryrzr.r|r2r2r3r}As $  zRandAugment._augmentation_spacer c Cs|j}t|\}}}t|tr*t|ttfrt|g|}n |dur*dd|D}||j||f}t |j D]B}tt t |d}t||} || \} } | jdkrbt| |jnd} | rpt ddrp| d9} t|| | |j|d }q8|S) img (PIL Image or Tensor): Image to be transformed. Returns: PIL Image or Tensor: Transformed image. NcSrr2rrr2r2r3rarz'RandAugment.forward..rrrr^rr)rr*rrrr.rr}rrangerrwrrrrkeysndimrr4r) rCr rrrrr_op_indexrrrrr2r2r3rTs"   zRandAugment.forwardc Cs:|jjd|jd|jd|jd|jd|jd }|S)Nz (num_ops=z , magnitude=z, num_magnitude_bins=, interpolation=rr)rEr5rrrrrrCsr2r2r3ros  zRandAugment.__repr__)r5r6r7r8rrr.rrrr@rrrrrr}rrrr2r2rDr3r s. .r c seZdZdZdejdfdededeee ddffdd Z d ede e e eefffd d Zd edefddZde fddZZS)r aDataset-independent data-augmentation with TrivialAugment Wide, as described in `"TrivialAugment: Tuning-free Yet State-of-the-Art Data Augmentation" `_. If the image is torch Tensor, it should be of type torch.uint8, and it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. If img is PIL Image, it is expected to be in mode "L" or "RGB". Args: num_magnitude_bins (int): The number of different magnitude values. interpolation (InterpolationMode): Desired interpolation enum defined by :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. fill (sequence or number, optional): Pixel fill value for the area outside the transformed image. If given a number, the value is used for all bands respectively. rNrrrr=cs t||_||_||_dSr>)r?r@rrr)rCrrrrDr2r3r@s  zTrivialAugmentWide.__init__rrcCstddftdd|dftdd|dftdd|dftdd|dftdd|dftdd|dftdd|dftdd|dftdd|dfdt||dd dftd d|dftddftddfd S) NrFgGz?Tg@@g`@rGrrRrvrr)rCrrr2r2r3r}s $  z&TrivialAugmentWide._augmentation_spacer c Cs|j}t|\}}}t|tr*t|ttfrt|g|}n |dur*dd|D}||j}tt t |d }t ||}||\} } | jdkr`t| t j t | dt jd nd} | rnt ddrn| d 9} t||| |j|d S) rNcSrr2rrr2r2r3rrz.TrivialAugmentWide.forward..rrdtyperr^rr)rr*rrrr.rr}rrwrrrrrrlongr4r) rCr rrrrrrrrrrr2r2r3rs$   $zTrivialAugmentWide.forwardcCs*|jjd|jd|jd|jd}|S)Nz(num_magnitude_bins=rrr)rEr5rrrrr2r2r3rs zTrivialAugmentWide.__repr__)r5r6r7r8rrr.rrrr@rrrrrr}rrrr2r2rDr3r |s" " r cseZdZdZdddddejdfdeded ed ed ed ed e e eddffdd Z dede eefde ee eefffddZejjdefddZejjdefddZdedefddZdedefddZdefd d!ZZS)"r aAugMix data augmentation method based on `"AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty" `_. If the image is torch Tensor, it should be of type torch.uint8, and it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions. If img is PIL Image, it is expected to be in mode "L" or "RGB". Args: severity (int): The severity of base augmentation operators. Default is ``3``. mixture_width (int): The number of augmentation chains. Default is ``3``. chain_depth (int): The depth of augmentation chains. A negative value denotes stochastic depth sampled from the interval [1, 3]. Default is ``-1``. alpha (float): The hyperparameter for the probability distributions. Default is ``1.0``. all_ops (bool): Use all operations (including brightness, contrast, color and sharpness). Default is ``True``. interpolation (InterpolationMode): Desired interpolation enum defined by :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.NEAREST``. If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.BILINEAR`` are supported. fill (sequence or number, optional): Pixel fill value for the area outside the transformed image. If given a number, the value is used for all bands respectively. rXrTNseverity mixture_width chain_depthalphaall_opsrrr=csntd|_d|kr|jks ntd|jd|d||_||_||_||_||_||_ ||_ dS)Nrrz!The severity must be between [1, z]. Got z instead.) r?r@_PARAMETER_MAXr1rrrrrrr)rCrrrrrrrrDr2r3r@s  zAugMix.__init__rrrsc Cstdd|dftdd|dftd|dd|dftd|dd|dftdd|dfdt||ddd ftd d|d ftdd ftdd fd }|jr|tdd |dftdd |dftdd |dftdd |dfd |S)NrrbTrg@rrurVFrv) rrrrrr#r$r%r&rc)rr r!r")rwrxryrzr.r{rupdate)rCrrrsrr2r2r3r}s&$   zAugMix._augmentation_spacecC t|Sr>)r*Z pil_to_tensorrCr r2r2r3_pil_to_tensor zAugMix._pil_to_tensorr cCrr>)r*Z to_pil_imagerr2r2r3_tensor_to_pilrzAugMix._tensor_to_pilparamscCrr>)rw_sample_dirichlet)rCrr2r2r3rrzAugMix._sample_dirichletorig_imgc Cst|j}t|\}}}t|tr-|}t|ttfr!t|g|}q2|dur,dd|D}n||}||j ||f}t |j }| dgt d|jd|} | dgdg| jd} |tj|j|jg| jd| dd} |tj|jg|j| jd| dd| dddf | ddg} | dddf | | } t|jD]x}| }|jdkr|jn ttjddd d }t|D]K}ttt|d }t ||}||\}}|jdkrt|tj|jd tjd nd }|rtd d r|d9}t||||j |d}q| !| dd|f | |q| |j"|j#d } t|ts8|$| S| S)rNcSrr2rrr2r2r3r/rz"AugMix.forward..rrVr)devicerr)lowhighsizerrr^rr)%rr*rrrr.rrr}rrshapeviewmaxrrrrwr{rrexpandrrrrrrrrrr4rZadd_torr)rCrrrrrr rZ orig_dimsbatchZ batch_dimsmZcombined_weightsZmixraugdepthrrrrrrr2r2r3r!sT    "$(  ""  zAugMix.forwardcCsJ|jjd|jd|jd|jd|jd|jd|jd|jd}|S) Nz (severity=z, mixture_width=z, chain_depth=z, alpha=z , all_ops=rrr) rEr5rrrrrrrrr2r2r3r[s"  zAugMix.__repr__)r5r6r7r8rZBILINEARr.rrrrr@rrrrr}rwZjitZunusedrrrrrrr2r2rDr3r sD  .:r )r+enumrtypingrrwrrr*r__all__rrrr4rnnModuler r r r r2r2r2r3s0     P 8]V