o wi#/@sHddlZddlZddlmZmZmZdZGdddZGdddZ dS)N)Image ImageEnhanceImageOps c@s*eZdZdZd ddZddZddZd S) ImageNetPolicyaDefinition of an ImageNetPolicy. Implements a fixed AutoAugment data augmentation policy targeted at ImageNet training by randomly applying at runtime one of the 25 pre-defined data augmentation sub-policies provided in Reference [1]. Usage example as a Pytorch Transform: >>> transform=transforms.Compose([transforms.Resize(256), >>> ImageNetPolicy(), >>> transforms.ToTensor()]) rrc Cstdddddd|tdddd dd|td d dd dd |tddd ddd|td dd ddd|td dddd d|tddd d dd |tdd dd dd|tddd ddd|td ddddd|tdd dddd|tdddd dd|td dd d d d|tdddd dd|tdddddd|tdd dddd|tdd ddd d |tddd ddd|tdddd dd|tdddd dd |td dd ddd|tdddd dd|tdddd dd|tdddddd|td d dd dd |g|_dS)zInitialize an ImageNetPolicy. Args: fillcolor (tuple): RGB color components of the color to be used for filling when needed (default: (128, 128, 128), which corresponds to gray). posterizeg?rotateg333333? solarize autocontrastequalizeg?g??colorrinvertcontrast sharpnessshearXN) SubPolicypolicies)self fillcolorr!n/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/nemo/collections/vision/data/megatron/autoaugment.py__init__%s4 zImageNetPolicy.__init__cCs$tdt|jd}|j||S)z,Define call method for ImageNetPolicy class.r)randomrandintlenr)rimg policy_idxr!r!r"__call__NszImageNetPolicy.__call__cCsdS)z,Define repr method for ImageNetPolicy class.rr!)rr!r!r"__repr__SszImageNetPolicy.__repr__N)r)__name__ __module__ __qualname____doc__r#r*r+r!r!r!r"rs  ) rc@s eZdZdZddZddZdS)rzDefinition of a SubPolicy. A SubPolicy consists of two augmentation operations, each of those parametrized as operation, probability, magnitude. The two operations are applied sequentially on the image upon call. c s.gd}||vr ||vsJdd|krdkr-nJdd|kr,dks1JdJdt|trDd|krCdksHJdJdt|tr[d|krZdks_Jd Jd td } tdd | tdd | tdd | tdd | tdd | tdd| ttdd| dtjtdd| tdd| tdd| tdd| dg| dg| dg| d} ddfddfddfddfddfddddddddddddddd dd!dd"dd} ||_| ||_ | |||_ ||_ | ||_ | |||_ d#S)$a+Initialize a SubPolicy. Args: operation1 (str): Key specifying the first augmentation operation. There are fourteen key values altogether (see supported_ops below listing supported operations). probability1 (float): Probability within [0., 1.] of applying the first augmentation operation. magnitude_idx1 (int): Integer specifiying the strength of the first operation as an index further used to derive the magnitude from a range of possible values. operation2 (str): Key specifying the second augmentation operation. probability2 (float): Probability within [0., 1.] of applying the second augmentation operation. magnitude_idx2 (int): Integer specifiying the strength of the second operation as an index further used to derive the magnitude from a range of possible values. fillcolor (tuple): RGB color components of the color to be used for filling. Returns: )rshearY translateX translateYr rr r rr brightnessrrrzCSubPolicy:one of oper1 or oper2 refers to an unsupported operation.rrz5SubPolicy: prob1 and prob2 should be within [0., 1.].rrzASubPolicy: idx1 should be specified as an integer within [0, 10].zASubPolicy: idx2 should be specified as an integer within [0, 10].r$g333333?gt?g?r rcSs6|d|}t|td|jd|}||jS)aDefine rotation transformation with fill. The input image is first rotated, then it is blended together with a gray mask of the same size. Note that fillcolor as defined elsewhere in this module doesn't apply here. Args: magnitude (float): rotation angle in degrees. Returns: rotated_filled (PIL Image): rotated image with gray filling for disoccluded areas unveiled by the rotation. RGBA)rrrr)convertr r compositenewsizemode)r( magnituderotatedrotated_filledr!r!r"rotate_with_fills  z,SubPolicy.__init__..rotate_with_fillc s4|j|jtjd|tddgddddftjdS)Nr$rr  transformr:rAFFINEr%choiceBICUBICr(r<rAr!r"z$SubPolicy.__init__..c s4|j|jtjddd|tddgddftjdSNr$rr@rArBrGrAr!r"rHrIc s:|j|jtjdd||jdtddgdddfdSrJrCr:rrDr%rErGrAr!r"rH &c s:|j|jtjddddd||jdtddgfdSrJrKrGrAr!r"rHrLcs ||SNr!rG)r?r!r"rH cS"t|d|tddgSNr$r@)rColorenhancer%rErGr!r!r"rHs"cS t||SrM)rr rGr!r!r"rH cSrSrM)rr rGr!r!r"rHrTcSrOrP)rContrastrRr%rErGr!r!r"rH cSrOrP)r SharpnessrRr%rErGr!r!r"rHrVcSrOrP)r BrightnessrRr%rErGr!r!r"rHrVcS t|SrM)rrrGr!r!r"rHrNcSrYrM)rrrGr!r!r"rHrNcSrYrM)rrrGr!r!r"rHrNN) isinstanceint _MAX_LEVELnplinspaceroundastypeint64 probability1 operation1 magnitude1 probability2 operation2 magnitude2) rrcrbmagnitude_idx1rfremagnitude_idx2r supported_ops num_levelsranges func_dictr!)r r?r"r#`s                /  zSubPolicy.__init__cCs<t|jkr|||j}t|jkr|||j}|S)z'Define call method for SubPolicy class.)r%rbrcrdrerfrg)rr(r!r!r"r*s zSubPolicy.__call__N)r,r-r.r/r#r*r!r!r!r"rXs  r) r%numpyr]PILrrrr\rrr!r!r!r"s @