keras: Data Load & Data Preprocessing_keras data dataloader

---

Data Load

Format Transformation

original format:

• Images
• Text files
• CSV data

you need to make your data available as one of 3 formats:

• NumPy arrays
  适合不大的数据
• tf.data.Dataset objects :
  ①有着GPU优化，比其他类型能更好地利用GPU。
  ②能从磁盘上读取大到内存放不下的数据。
• Python generators

转化为tf.data.Dataset

读取

• Images: tf.keras.preprocessing.image_dataset_from_directory(...)

# image files sorted into class-specific folders
main_directory/
...class_a/
......a_image_1.jpg
......a_image_2.jpg
...class_b/
......b_image_1.jpg
......b_image_2.jpg
1
2
3
4
5
6
7
8

dataset = keras.preprocessing.image_dataset_from_directory(
  'path/to/main_directory', batch_size=64, image_size=(200, 200))

# For demonstration, iterate over the batches yielded by the dataset.
for data, labels in dataset:
   print(data.shape)  # (64, 200, 200, 3) 每批64张、200*200像素、3个RGB通道
   print(data.dtype)  # float32
   print(labels.shape)  # (64,)	每批标签64个
   print(labels.dtype)  # int32
1
2
3
4
5
6
7
8
9

• Text files: keras.preprocessing.text_dataset_from_directory(...)
  同样，在不同文件夹中按类分类的文档。

dataset = keras.preprocessing.text_dataset_from_directory(
  'path/to/main_directory', batch_size=64)

# For demonstration, iterate over the batches yielded by the dataset.
for data, labels in dataset:
   print(data.shape)  # (64,)
   print(data.dtype)  # string
   print(labels.shape)  # (64,)
   print(labels.dtype)  # int32
1
2
3
4
5
6
7
8
9

• other:
  tf.data.experimental.make_csv_dataset to load structured data from CSV files.
  tf.data.Dataset.from_tensor_slices()：keras: tf.data.Dataset.from_tensor_slices()

操作

• 查看方式1：迭代.as_numpy_iterator()

print(list(dataset.as_numpy_iterator()))
# [(array([1, 3], dtype=int32), array([b'A'], dtype=object)), 
#  (array([2, 1], dtype=int32), array([b'B'], dtype=object)), 
#  (array([3, 3], dtype=int32), array([b'A'], dtype=object))]
1
2
3
4

• 查看方式2：for

for element in dataset.as_numpy_iterator():
	print(element)
# (array([1, 3], dtype=int32), array([b'A'], dtype=object))
# (array([2, 1], dtype=int32), array([b'B'], dtype=object))
# (array([3, 3], dtype=int32), array([b'A'], dtype=object))
1
2
3
4
5

• .take(count)：取出几批的样本。

for inputs, targets in dataset.take(1):
	print(inputs)			# tf.Tensor([1 3], shape=(2,), dtype=int32)
	print(targets)			# tf.Tensor([b'A'], shape=(1,), dtype=string)
1
2
3

• .batch()：指定batch_size。必须指定，不然fit()时会报错。

# 指定一批32个
dataset = dataset.batch(32)
1
2

Data Preprocessing

vectorized & standardized

简单来说：

• vectorized 向量化：非数字特征映射到数字，比如[狗, 猫]→[0, 1]
• standardized 标准化：修改范围到[0.0, 1.0]、符合概率学（均值0和方差1）

详细：

• Text files
  ①need to be read into string tensors,
  ②then split into words.
  ③Finally, the words need to be indexed & turned into integer tensors.
• Images
  ①need to be read and decoded into integer tensors,
  ②then converted to floating point and normalized to small values (usually between 0 and 1).
• CSV data
  ①needs to be parsed, with numerical features converted to floating point tensors and categorical features indexed and converted to integer tensors.
  ②Then each feature typically needs to be normalized to zero-mean and unit-variance.

Text

基本

tensorflow.keras.layers.experimental.preprocessing.TextVectorization：holds a mapping between string tokens and integer indices.

• 词汇表必须是字符串。
• 索引0表示缺省值(即单词长度不够时的空单词"")，索引1表示词汇表外的值(词汇表由adapt()指定)。

from tensorflow.keras.layers.experimental import preprocessing

vocabulary = ["aa bb cc"]
data = ["aa bb cc"]
layer = preprocessing.TextVectorization()
layer.adapt(vocabulary)						# 以哪个为词汇表
normalized_data = layer(data)				# 根据之前adapt()的vocabulary翻译data
print(normalized_data)
# tf.Tensor([[4 3 2 2 1 1]], shape=(1, 6), dtype=int64)
1
2
3
4
5
6
7
8
9

• 重复的单词cc，可以看到都是2。
• 词汇表外的值dd和ee，都是1。
• 词汇表vocabulary映射adapt()时，标点符号和空格不算，只看单词。重复的单词只留一个。
• 词汇表vocabulary可以是一维数组["aa bb cc"](句子)、["aa bb", "bb cc"](句子)、["aa", "bb", "cc"](单词)，不能是字符串"aa bb cc"，不能是多列二维数组[["aa", "bb"], ["aa", "cc"]]，但可以是单列的二维数组[["aa bb"], ["aa cc"]](句子)、[["aa"], ["bb"], ["cc"]](单词)。
• 处理data同样也是同样的格式要求，结果的形状必定是二维。注意，认为每行是一个"..."。

data = ["aa bb cc", "cc dd"]
'''
tf.Tensor(
[[2 4 3]
 [3 1 0]], shape=(2, 3), dtype=int64)
'''
1
2
3
4
5
6

• 单词长度不够，指的是data中的两句话，选取最长单词数作为结果的列维度，其他不足长度的句子少的单词就对应0。

热编码 one-hot encoded

# Example: one-hot encoded bigrams
from tensorflow.keras.layers.experimental import preprocessing

vocabulary = ["aa bb cc"]
data = ["aa", "bb", "cc", "dd", ""]

layer = preprocessing.TextVectorization(output_mode="binary", ngrams=2)
layer.adapt(vocabulary)

integer_data = layer(data)
print(integer_data)
'''
tf.Tensor(
[[0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 1. 0. 0.]
 [0. 1. 0. 0. 0. 0.]
 [1. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0.]], shape=(5, 6), dtype=float32)
'''
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

每行都只有一个位是1，其他都是0.

Image & CSV: normalizing features

• 均值0和方差1：tensorflow.keras.layers.experimental.preprocessing.Normalization
  adapt()接收三类输入类型：a batched Dataset, a Tensor, or a Numpy array。不能直接用pd.DataFrame.

from tensorflow.keras.layers.experimental import preprocessing

data = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=np.float32)

normalizer = preprocessing.Normalization()
normalizer.adapt(data)

normalized_data = normalizer(data)
print(normalized_data)
'''
tf.Tensor(
[[-1.2247448 -1.2247448 -1.2247448]
 [ 0.         0.         0.       ]
 [ 1.2247448  1.2247448  1.2247448]], shape=(3, 3), dtype=float32)
'''
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

# tf.keras.utils.normalize(): numpy array
normalized_data = tf.keras.utils.normalize(data)
print(normalized_data)
1
2
3

• 调整范围：tensorflow.keras.layers.experimental.preprocessing.Rescaling

import numpy as np
from tensorflow.keras.layers.experimental import preprocessing

# Example image data, with values in the [0, 255] range
training_data = np.random.randint(0, 256, size=(64, 200, 200, 3)).astype("float32")

# 限定范围：从[0, 255]到[0.0, 1.0]
output_data = preprocessing.Rescaling(scale=1.0 / 255)(training_data)
1
2
3
4
5
6
7
8

如果是numpy，那么可以直接

a = np.array([25.5,255])
a = a/255
1
2

labels分类

• num_classes（这里是3）必须大于等于labels的最大值+1.
• y表示的类别应该是[0, MAX]，这样恰好符合num_classes。如果从1开始的话，虽然可以，但是创出来就是有一个从没有用到的0列。

y = np.array([0, 2, 1, 2, 1]);		# 三类：0 1 2
y = keras.utils.to_categorical(y, 3)
print(y)
'''
[[1. 0. 0.]
 [0. 0. 1.]
 [0. 1. 0.]
 [0. 0. 1.]
 [0. 1. 0.]]
'''
1
2
3
4
5
6
7
8
9
10

更多

为了处理

Categorical data preprocessing layers

• CategoryEncoding layer
• Hashing layer
• Discretization layer
• StringLookup layer
• IntegerLookup layer
• CategoryCrossing layer

Image preprocessing & augmentation layers

• Resizing layer
• Rescaling layer
• CenterCrop layer
• RandomCrop layer
• RandomFlip layer
• RandomTranslation layer
• RandomRotation layer
• RandomZoom layer
• RandomHeight layer
• RandomWidth layer

Core preprocessing layers

• TextVectorization layer
• Normalization layer

为了生成Dataset

Dataset preprocessing

• Image data preprocessing
  • image_dataset_from_directory function
  • load_img function
  • img_to_array function
  • ImageDataGenerator class
  • flow method
  • flow_from_dataframe method
  • flow_from_directory method
• Timeseries data preprocessing
  • timeseries_dataset_from_array function
  • pad_sequences function
  • TimeseriesGenerator class
• Text data preprocessing
  • text_dataset_from_directory function
  • Tokenizer class

数据处理层可以写入到Model中

normalizer = preprocessing.Normalization()
normalizer.adapt(x_train)

inputs = keras.Input(shape=input_shape)
x = preprocessing_layer(inputs)
outputs = rest_of_the_model(x)
model = keras.Model(inputs, outputs)
1
2
3
4
5
6
7