BaseTransform

    视觉中图像变化的基类。

    调用逻辑:

    如果你想要定义自己的图像变化方法, 需要重写子类中的 _apply_* 方法。

    1. import numpy as np
    2. from PIL import Image
    3. import paddle.vision.transforms.functional as F
    4. from paddle.vision.transforms import BaseTransform
    6. def _get_image_size(img):
    7.     if F._is_pil_image(img):
    8.         return img.size
    9.     elif F._is_numpy_image(img):
    10.         return img.shape[:2][::-1]
    11.     else:
    12.         raise TypeError("Unexpected type {}".format(type(img)))
    14. class CustomRandomFlip(BaseTransform):
    15.     def __init__(self, prob=0.5, keys=None):
    16.         super(CustomRandomFlip, self).__init__(keys)
    18.     def _get_params(self, inputs):
    19.         image = inputs[self.keys.index('image')]
    20.         params = {}
    21.         params['flip'] = np.random.random() < self.prob
    22.         params['size'] = _get_image_size(image)
    23.         return params
    25.     def _apply_image(self, image):
    26.         if self.params['flip']:
    27.             return F.hflip(image)
    28.         return image
    30.     # if you only want to transform image, do not need to rewrite this function
    31.     def _apply_coords(self, coords):
    32.         if self.params['flip']:
    33.             w = self.params['size'][0]
    34.             coords[:, 0] = w - coords[:, 0]
    35.         return coords
    37.     # if you only want to transform image, do not need to rewrite this function
    39.         idxs = np.array([(0, 1), (2, 1), (0, 3), (2, 3)]).flatten()
    40.         coords = np.asarray(boxes).reshape(-1, 4)[:, idxs].reshape(-1, 2)
    41.         minxy = coords.min(axis=1)
    42.         maxxy = coords.max(axis=1)
    43.         trans_boxes = np.concatenate((minxy, maxxy), axis=1)
    44.         return trans_boxes
    46.     # if you only want to transform image, do not need to rewrite this function
    47.     def _apply_mask(self, mask):
    48.         if self.params['flip']:
    49.             return F.hflip(mask)
    50.         return mask
    52. # create fake inputs
    53. fake_img = Image.fromarray((np.random.rand(400, 500, 3) * 255.).astype('uint8'))
    54. fake_boxes = np.array([[2, 3, 200, 300], [50, 60, 80, 100]])
    55. fake_mask = fake_img.convert('L')
    57. # only transform for image:
    58. flip_transform = CustomRandomFlip(1.0)
    59. converted_img = flip_transform(fake_img)
    61. # transform for image, boxes and mask
    62. flip_transform = CustomRandomFlip(1.0, keys=('image', 'boxes', 'mask'))
    64. print('converted boxes', converted_boxes)