import numpy as np from PIL import Image from typing import (TYPE_CHECKING, Annotated, Any, Dict, List, Optional, Tuple, Union) from pydantic import BaseModel, Field, model_validator class ResizeImage(BaseModel): size: int | tuple[int, int] = (256, 480) # H,W def __call__(self): try: from torchvision.transforms import v2 except: from torchvision import transforms as v2 return v2.Resize(self.size, interpolation=v2.InterpolationMode.NEAREST) @property def resolution(self) -> Tuple[int, int]: if isinstance(self.size, int): return (self.size, self.size) elif isinstance(self.size, list) and len(self.size) == 2: return (self.size[0], self.size[1]) elif isinstance(self.size, tuple) and len(self.size) == 2: return self.size else: raise ValueError("size should be int or list of two ints") class ColorJitter(BaseModel): brightness: float | tuple[float, float] = Field(default_factory=lambda: 0.2) contrast: float | tuple[float, float] = Field(default_factory=lambda: (0.8, 1.2)) saturation: float | tuple[float, float] = Field(default_factory=lambda: (0.8, 1.2)) hue: float | tuple[float, float] = Field(default_factory=lambda: 0.05) def __call__(self): try: from torchvision.transforms import v2 except: from torchvision import transforms as v2 return v2.ColorJitter( brightness=self.brightness, contrast=self.contrast, saturation=self.saturation, hue=self.hue, ) class CenterCrop(BaseModel): size: int | tuple[int, int] = (224,224) # H,W def __call__(self): try: from torchvision.transforms import v2 except: from torchvision import transforms as v2 return v2.CenterCrop(self.size) class Normalize(BaseModel): mean: float | List[float] = Field(default_factory=lambda: [0.485, 0.456, 0.406]) std: float | List[float] = Field(default_factory=lambda: [0.229, 0.224, 0.225]) def __call__(self): try: from torchvision.transforms import v2 except: from torchvision import transforms as v2 mean = [self.mean] * 3 if isinstance(self.mean, float) else self.mean std = [self.std] * 3 if isinstance(self.std, float) else self.std return v2.Normalize(mean, std) # type: ignore class GaussianNoise(BaseModel): mean: float = 0 std: float = 3 prob_skip: float = 0.1 def __call__(self, img: Image.Image): rnd = np.random.rand() if rnd < self.prob_skip: return img # 10% chance to skip adding noise arr = np.array(img).astype(np.float32) # mean = int(np.random.uniform(-self.mean, self.mean)) # real mean noise = np.random.normal(self.mean, self.std, arr.shape) arr_noisy = arr + noise arr_noisy = np.clip(arr_noisy, 0, 255).astype(np.uint8) return Image.fromarray(arr_noisy) class SaltPepperNoise: """ Adds salt-and-pepper noise to a tensor image. - prob: total probability of altering a pixel - salt_vs_pepper: fraction of salt (white) vs pepper (black) """ def __init__(self, prob=0.01, salt_vs_pepper=0.5): self.prob = prob self.salt_vs_pepper = salt_vs_pepper def __call__(self, img): # img is PIL Image arr = np.array(img) # H x W x C # Random mask rnd = np.random.rand(arr.shape[0], arr.shape[1]) pepper_mask = rnd < (self.prob * (1 - self.salt_vs_pepper)) salt_mask = rnd > (1 - self.prob * self.salt_vs_pepper) arr = arr.copy() # Pepper = 0 arr[pepper_mask] = 0 # Salt = 255 arr[salt_mask] = 255 return Image.fromarray(arr) def __repr__(self): return f"{self.__class__.__name__}(prob={self.prob}, salt_vs_pepper={self.salt_vs_pepper})"