tensorflow数据加载与前处理

一、数据加载

1、加载文件夹下图片数据：不同类图片放在不同的文件夹下

batch_size = 32　　#next(iter(train_ds))，一次迭代32张图片
img_height = 224
img_width = 224
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
　　str(data_root), validation_split=0.2, subset="training", seed=123, image_size=(img_height, img_width), batch_size=batch_size　　#data_root是分类图片路径
)　　
print(train_ds.class_names)　　#图片类别/文件夹个数

2、加载tensorflow_datasets数据：自动将数据分成训练数据、验证数据、测试数据

import tensorflow_datasets as tfds
(train_ds, val_ds, test_ds), metadata = tfds.load(
    'tf_flowers',
    split=['train[:80%]', 'train[80%:90%]', 'train[90%:]'],
    with_info=True,
    as_supervised=True,
)
print(metadata.features['label'].num_classes)　　#dataset有几个类别
get_label_name = metadata.features['label'].int2str
get_label_name(0)　　#label对应的类别

 二、数据前处理

https://tensorflow.google.cn/tutorials/images/data_augmentation?hl=zh_cn

1、用lambda表达式对整个数据集进行处理（5中也涉及）

normalization_layer = tf.keras.layers.experimental.preprocessing.Rescaling(1./255)　　#数据缩放到[0, 1]
train_ds = train_ds.map(lambda x, y: (normalization_layer(x), y))

 2、通用数据增强层

1）layers.Lambda层进行图像增强

def random_invert_img(x, p=0.5):
    if  tf.random.uniform([]) < p:
        x = (255-x)
    else:
        x
    return x

def random_invert(factor=0.5):
  return layers.Lambda(lambda x: random_invert_img(x, factor))

random_invert(image)　　#layers.Lambda会自动调用绑定的函数处理图片的每个像素

2）通过subclassing新建一个层

class RandomInvert(layers.Layer):
  def __init__(self, factor=0.5, **kwargs):
    super().__init__(**kwargs)
    self.factor = factor
  def call(self, x):
    return random_invert_img(x, 0.3)　　#random_invert_img()函数上面有

_ = plt.imshow(RandomInvert()(image)[0])

3、tf.image、tf.data实现数据增强管道或层

flipped = tf.image.flip_left_right(image)
grayscaled = tf.image.rgb_to_grayscale(image); tf.squeeze(grayscaled)
saturated = tf.image.adjust_saturation(image, 3)
bright = tf.image.adjust_brightness(image, 0.4)
cropped = tf.image.central_crop(image, central_fraction= 0.5)
rotated = tf.image.rot90(image)

4、随便变换（）

tf.image.stateless_random_brightness
tf.image.stateless_random_contrast
tf.image.stateless_random_crop
tf.image.stateless_random_flip_left_right
tf.image.stateless_random_flip_up_down
tf.image.stateless_random_hue
tf.image.stateless_random_jpeg_quality
tf.image.stateless_random_saturation
例如：
seed = (i, 0)  # tuple of size (2,)；相同种子，同样结果
stateless_random_brightness = tf.image.stateless_random_brightness(
      image, max_delta=0.95, seed=seed)    #随机改变图片亮度
 stateless_random_contrast = tf.image.stateless_random_contrast(
      image, lower=0.1, upper=0.9, seed=seed)    #随机改变图片对比度
stateless_random_crop = tf.image.stateless_random_crop(
      image, size=[210, 300, 3], seed=seed)    #随机扣取图片

5、数据集增强

#下载数据集
(train_datasets, val_ds, test_ds), metadata = tfds.load(
    'tf_flowers',
    split=['train[:80%]', 'train[80%:90%]', 'train[90%:]'],
    with_info=True,
    as_supervised=True,
)
#图像缩放
def resize_and_rescale(image, label):
  image = tf.cast(image, tf.float32)
  image = tf.image.resize(image, [IMG_SIZE, IMG_SIZE])
  image = (image / 255.0)
  return image, label
#带seed的增强函数
def augment(image_label, seed):
  image, label = image_label
  image, label = resize_and_rescale(image, label)
  image = tf.image.resize_with_crop_or_pad(image, IMG_SIZE + 6, IMG_SIZE + 6)
  # Make a new seed
  new_seed = tf.random.experimental.stateless_split(seed, num=1)[0, :]
  # Random crop back to the original size
  image = tf.image.stateless_random_crop(
      image, size=[IMG_SIZE, IMG_SIZE, 3], seed=seed)
  # Random brightness
  image = tf.image.stateless_random_brightness(
      image, max_delta=0.5, seed=new_seed)
  image = tf.clip_by_value(image, 0, 1)
  return image, label

1）利用tf.data.experimental.Counter()

（1）？创建一个计数器，并使用（counter, counter）压缩数据集，这将确保数据集中的每个图像斗鱼基于计数器的唯一值相关联

Create a tf.data.experimental.Counter() object (let's call it counter) and zip the dataset with (counter, counter). This will ensure that each image in the dataset gets associated with a unique value (of shape (2,)) based on counter which later can get passed into the augment function as the seed value for random transformations.

counter = tf.data.experimental.Counter()
train_ds = tf.data.Dataset.zip((train_datasets, (counter, counter)))

(2)利用映射函数处理数据集

train_ds = (
    train_ds
    .shuffle(1000)
    .map(augment, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)
val_ds = (
    val_ds
    .map(resize_and_rescale, num_parallel_calls=AUTOTUNE)
    .batch(batch_size)
    .prefetch(AUTOTUNE)
)
test_ds = ......

2）利用tf.random.Generator

（1）创建映射函数

Note: tf.random.Generator objects store RNG state in a tf.Variable, which means it can be saved as a checkpoint or in a SavedModel. For more details, please refer to Random number generation.

rng = tf.random.Generator.from_seed(123, alg='philox')    # Create a generator
#映射函数，映射过程如上一致
def f(x, y):
  seed = rng.make_seeds(2)[0]
  image, label = augment((x, y), seed)
  return image, label    # A wrapper function for updating seeds