使用数据集 api 在 TensorFlow 中进行数据扩充的正确方法？

Question

所以，我一直在使用 TensorFlow 数据集 API 来加载图像和分割蒙版（用于语义分割项目），我希望能够生成批量图像和蒙版，每张图像都随机经历了亮度变化、对比度变化、裁剪、饱和度变化等预处理功能的任意组合。因此，我的批次中的第一张图像可能没有预处理，第二张可能有饱和度变化，第三张可能有亮度和饱和度等等。

我尝试了以下方法：

import tensorflow as tf
from tensorflow.contrib.data import Dataset, Iterator
import random


def _resize_image(image, mask):
    image = tf.image.resize_bicubic(image, [480, 640], True)
    mask = tf.image.resize_bicubic(mask, [480, 640], True)
    return image, mask

def _corrupt_contrast(image, mask):
    image = tf.image.random_contrast(image, 0, 5)
    return image, mask


def _corrupt_saturation(image, mask):
    image = tf.image.random_saturation(image, 0, 5)
    return image, mask


def _corrupt_brightness(image, mask):
    image = tf.image.random_brightness(image, 5)
    return image, mask


def _random_crop(image, mask):
    seed = random.random()
    image = tf.random_crop(image, [240, 320, 3], seed=seed)
    mask = tf.random_crop(mask, [240, 320, 1], seed=seed)
    return image, mask


def _flip_image_horizontally(image, mask):
    seed = random.random()
    image = tf.image.random_flip_left_right(image, seed=seed)
    mask = tf.image.random_flip_left_right(mask, seed=seed)

    return image, mask


def _flip_image_vertically(image, mask):
    seed = random.random()
    image = tf.image.random_flip_up_down(image, seed=seed)
    mask = tf.image.random_flip_up_down(mask, seed=seed)

    return image, mask


def _normalize_data(image, mask):
    image = tf.cast(image, tf.float32)
    image = image / 255.0

    mask = tf.cast(mask, tf.float32)
    mask = mask / 255.0

    return image, mask


def _parse_data(image_paths, mask_paths):
    image_content = tf.read_file(image_paths)
    mask_content = tf.read_file(mask_paths)

    images = tf.image.decode_png(image_content, channels=3)
    masks = tf.image.decode_png(mask_content, channels=1)

    return images, masks


def data_batch(image_paths, mask_paths, params, batch_size=4, num_threads=2):
    # Convert lists of paths to tensors for tensorflow
    images_name_tensor = tf.constant(image_paths)
    mask_name_tensor = tf.constant(mask_paths)

    # Create dataset out of the 2 files:
    data = Dataset.from_tensor_slices(
        (images_name_tensor, mask_name_tensor))

    # Parse images and labels
    data = data.map(
        _parse_data, num_threads=num_threads, output_buffer_size=6 * batch_size)

    # Normalize images and masks for vals. between 0 and 1
    data = data.map(_normalize_data, num_threads=num_threads, output_buffer_size=6 * batch_size)

    if params['crop'] and not random.randint(0, 1):
        data = data.map(_random_crop, num_threads=num_threads,
                    output_buffer_size=6 * batch_size)

    if params['brightness'] and not random.randint(0, 1):
        data = data.map(_corrupt_brightness, num_threads=num_threads,
                    output_buffer_size=6 * batch_size)

    if params['contrast'] and not random.randint(0, 1):
        data = data.map(_corrupt_contrast, num_threads=num_threads,
                    output_buffer_size=6 * batch_size)

    if params['saturation'] and not random.randint(0, 1):
        data = data.map(_corrupt_saturation, num_threads=num_threads,
                    output_buffer_size=6 * batch_size)

    if params['flip_horizontally'] and not random.randint(0, 1):
        data = data.map(_flip_image_horizontally,
                    num_threads=num_threads, output_buffer_size=6 * batch_size)

    if params['flip_vertically'] and not random.randint(0, 1):
        data = data.map(_flip_image_vertically, num_threads=num_threads,
                    output_buffer_size=6 * batch_size)

    # Shuffle the data queue
    data = data.shuffle(len(image_paths))

    # Create a batch of data
    data = data.batch(batch_size)

    data = data.map(_resize_image, num_threads=num_threads,
                    output_buffer_size=6 * batch_size)

    # Create iterator
    iterator = Iterator.from_structure(data.output_types, data.output_shapes)

    # Next element Op
    next_element = iterator.get_next()

    # Data set init. op
    init_op = iterator.make_initializer(data)

    return next_element, init_op

但是由此返回的所有批次都应用了相同的转换，而不是不同的组合，我的猜测是 random.randint 持续存在，并且实际上不是每个批次的运行，如果是这样的话，我该如何解决这个问题以获得预期的结果？有关我打算如何使用它的示例（我觉得这与问题无关，但人们可能仍然想知道）可以找到 here

Answer 1

所以问题确实是带有 if 语句的控制流带有 Python 变量，并且只在创建图形时执行一次，为了做我想做的事，我必须定义一个包含是否应用函数的布尔值的占位符（并在每次迭代中输入一个新的布尔张量以更改扩充），并且控制流由 tf.cond 处理。如果有人感兴趣，我将新代码推送到 GitHub link 我在上面的问题中发布。

使用数据集 api 在 TensorFlow 中进行数据扩充的正确方法？

Correct way of doing data augmentation in TensorFlow with the dataset api?

python-3.x

tensorflow

tensorflow-datasets