【tensorflow2.0】数据管道dataset

 从Python generator构建数据管道
from matplotlib  pyplot as plt 
from tensorflow.keras.preprocessing.image  ImageDataGenerator
 
 定义一个从文件中读取图片的generator
image_generator = ImageDataGenerator(rescale=1.0/255).flow_from_directory(
                    ./data/cifar2/test/),batch_size=20binary)
 
classdict = image_generator.class_indices
print(classdict)

def generator():
    in image_generator:
        yield (features,label)
 
ds3 = tf.data.Dataset.from_generator(generator,output_types=(tf.float32,tf.int32))
%matplotlib inline
%config InlineBackend.figure_format = svg
plt.figure(figsize=(6,6)) 
for i,(img,label) in enumerate(ds3.unbatch().take(9)):
    ax=plt.subplot(3,3,i+1)
    ax.imshow(img.numpy())
    ax.set_title(label = %d"%label)
    ax.set_xticks([])
    ax.set_yticks([]) 
plt.show()

 从csv文件构建数据管道
ds4 = tf.data.experimental.make_csv_dataset(
      file_pattern = [./data/titanic/train.csv",1)">./data/titanic/test.csv=Survived""=True)
 
for data,1)">in ds4.take(2print(data,label)

 从文本文件构建数据管道
 
ds5 = tf.data.TextLineDataset(
    filenames = []
    ).skip(1) 略去第一行header
 
for line in ds5.take(5print(line)

s6 = tf.data.Dataset.list_files(./data/cifar2/train/*/*.jpg)
for file in ds6.take(5print(file)

def load_image(img_path,size = (32,1)">)):
    label = 1 if tf.strings.regex_full_match(img_path,1)">.*/automobile/.*") else 0
    img = tf.io.read_file(img_path)
    img = tf.image.decode_jpeg(img) 注意此处为jpeg格式
    img = tf.image.resize(img,size)
    return(img,label)
 
%'
in enumerate(ds6.map(load_image).take(2)):
    plt.figure(i)
    plt.imshow((img/255.0).numpy())
    plt.title(label)
    plt.xticks([])
    plt.yticks([])

 os
 numpy as np
 
 inpath：原始数据路径 outpath:TFRecord文件输出路径
 create_tfrecords(inpath,outpath): 
    writer = tf.io.TFRecordWriter(outpath)
    dirs = os.listdir(inpath)
    for index,name  enumerate(dirs):
        class_path = inpath +/"+ name+"
        for img_name  os.listdir(class_path):
            img_path = class_path + img_name
            img = tf.io.read_file(img_path)
            img = tf.image.decode_image(img)
            img = tf.image.encode_jpeg(img) #统一成jpeg格式压缩
            example = tf.train.Example(
               features=tf.train.Features(feature={
                    label': tf.train.Feature(int64_list=tf.train.Int64List(value=[index])),1)">img_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[img.numpy()]))
               }))
            writer.write(example.SerializeToString())
    writer.close()
 
create_tfrecords(./data/cifar2_test.tfrecords/)
 
 pyplot as plt 
 
 parse_example(proto):
    description ={  : tf.io.FixedLenFeature([],tf.string),1)">: tf.io.FixedLenFeature([],tf.int64)} 
    example = tf.io.parse_single_example(proto,description)
    img = tf.image.decode_jpeg(example["])   注意此处为jpeg格式
    img = tf.image.resize(img,(32,1)">))
    label = example[]
    ./data/cifar2_test.tfrecords").map(parse_example).shuffle(3000)
 
%in enumerate(ds7.take(9)
    ax.imshow((img/255.0).numpy())
    ax.set_title(label)
    ax.set_xticks([])
    ax.set_yticks([]) 
plt.show()

 map:将转换函数映射到数据集每一个元素
 
ds = tf.data.Dataset.from_tensor_slices([hello worldhello Chinahello Beijing])
ds_map = ds.map(lambda x:tf.strings.split(x,1)">" ))
for x  ds_map:
    print(x)

 flat_map:将转换函数映射到数据集的每一个元素，并将嵌套的Dataset压平。
])
ds_flatmap = ds.flat_map(lambda x:tf.data.Dataset.from_tensor_slices(tf.strings.split(x,1)">)))
 ds_flatmap:
    print(x)

 interleave: 效果类似flat_map,但可以将不同来源的数据夹在一起。
])
ds_interleave = ds.interleave( ds_interleave:
     filter:过滤掉某些元素。
])
 找出含有字母a或B的元素
ds_filter = ds.filter(lambda x: tf.strings.regex_full_match(x,1)">.*[a|B].* ds_filter:
    print(x)

 zip:将两个长度相同的Dataset横向铰合。
 
ds1 = tf.data.Dataset.range(0,3)
ds2 = tf.data.Dataset.range(3,1)">)
ds3 = tf.data.Dataset.range(6,9)
ds_zip = tf.data.Dataset.zip((ds1,ds2,ds3))
for x,y,z  ds_zip:
    print(x.numpy(),y.numpy(),z.numpy())

 condatenate:将两个Dataset纵向连接。
)
ds_concat = tf.data.Dataset.concatenate(ds1,ds2)
 ds_concat:
    print(x)

 reduce:执行归并操作。
 
ds = tf.data.Dataset.from_tensor_slices([1,5.0])
result = ds.reduce(0.0,lambda x,y:tf.add(x,y))
result

 batch:构建批次，每次放一个批次。比原始数据增加一个维度。 其逆操作为unbatch。 
 
ds = tf.data.Dataset.range(12)
ds_batch = ds.batch(4 ds_batch:
    print(x)

 padded_batch:构建批次，类似batch,但可以填充到相同的形状。
 
elements = [[1,2],[3,5],[6,7],[8]]
ds = tf.data.Dataset.from_generator(: iter(elements),tf.int32)
 
ds_padded_batch = ds.padded_batch(2,padded_shapes = [4 ds_padded_batch:
    print(x)

 window:构建滑动窗口，返回Dataset of Dataset.
 window返回的是Dataset of Dataset,可以用flat_map压平
ds_window = ds.window(3,shift=1).flat_map(lambda x: x.batch(3,drop_remainder=True)) 
 ds_window:
    print(x)

 shuffle:数据顺序洗牌。
)
ds_shuffle = ds.shuffle(buffer_size = 5 ds_shuffle:
    print(x)

 repeat:重复数据若干次，不带参数时，重复无数次。
 
ds = tf.data.Dataset.range(3)
ds_repeat = ds.repeat(3 ds_repeat:
    
 shard:采样，从某个位置开始隔固定距离采样一个元素。
)
ds_shard = ds.shard(3,index = 1 ds_shard:
    
 take:采样，从开始位置取前几个元素。
)
ds_take = ds.take(3)
 
list(ds_take.as_numpy_iterator())

 tensorflow as tf
 
 打印时间分割线
@tf.function
 printbar():
    ts = tf.timestamp()
    today_ts = ts%(24*60*60)
 
    hour = tf.cast(today_ts//3600+8,tf.int32)%tf.constant(24)
    minite = tf.cast((today_ts%3600)//60 timeformat(m):
        if tf.strings.length(tf.strings.format({}:
            return(tf.strings.format(0{}= tf.strings.join([timeformat(hour),timeformat(minite),timeformat(second)],separator = :)
    tf.print(=========="*8,end = print(timestring)
 
 time
 
 数据准备和参数迭代两个过程默认情况下是串行的。
 
 模拟数据准备
for i in range(10):
        假设每次准备数据需要2s
        time.sleep(2) 
         i 
ds = tf.data.Dataset.from_generator(generator,output_types = (tf.int32))
 
 模拟参数迭代
 train_step():
    假设每一步训练需要1s
    time.sleep(1) 
 
 训练过程预计耗时 10*2+10*1+ = 30s
printbar()
tf.print(tf.constant(start training... ds:
    train_step()  
printbar()
tf.end training... 使用 prefetch 方法让数据准备和参数迭代两个过程相互并行。
 
 训练过程预计耗时 max(10*2,10*1) = 20s
start training with prefetch...))
 
 tf.data.experimental.AUTOTUNE 可以让程序自动选择合适的参数
in ds.prefetch(buffer_size = tf.data.experimental.AUTOTUNE):
    train_step()  
 
printbar()
tf."))

================================================================================16:18:31
start training...
================================================================================16:19:01
end training...
================================================================================16:19:01
start training with prefetch...
================================================================================16:19:23
end training...

ds_files = tf.data.Dataset.list_files(./data/titanic/*.csv)
ds = ds_files.flat_map(lambda x:tf.data.TextLineDataset(x).skip(1in ds.take(4
ds_files = tf.data.Dataset.list_files()
ds = ds_files.interleave(in ds.take(8print(line)

ds = tf.data.Dataset.list_files( 单进程转换
start transformation...))
 
ds_map = ds.map(load_image)
for _ pass
 
printbar()
tf.end transformation... 多进程转换
start parallel transformation...))
 
ds_map_parallel = ds.map(load_image,num_parallel_calls = tf.data.experimental.AUTOTUNE)
 ds_map_parallel:
    end parallel transformation..."))

in range(5假设每一步训练需要0s
    pass
 
 训练过程预计耗时 (5*2+5*0)*3 = 30s
for epoch in tf.range(3 ds:
        train_step()  
    printbar()
    tf.epoch = ended)
printbar()
tf. i 
 
 使用 cache 方法让数据在第一个epoch后缓存到内存中，仅限于数据集不大情形。
ds = tf.data.Dataset.from_generator(generator,1)"> (tf.int32)).cache()
 
假设每一步训练需要0s
    time.sleep(0) 
 
 训练过程预计耗时 (5*2+5*0)+(5*0+5*0)*2 = 10s
"))

================================================================================17:07:57
start training...
================================================================================17:08:07
epoch = 0  ended
================================================================================17:08:17
epoch = 1  ended
================================================================================17:08:27
epoch = 2
end training...
================================================================================17:08:27
start training...
================================================================================17:08:37 0  ended
================================================================================17:08:37  ended
================================================================================17:08:37
end training...

 先map后batch
ds = tf.data.Dataset.range(100000)
ds_map_batch = ds.map(lambda x:x**2).batch(20)
 
printbar()
tf.start scalar transformation... ds_map_batch:
    
printbar()
tf.end scalar transformation... 先batch后map
ds = tf.data.Dataset.range(100000)
ds_batch_map = ds.batch(20).map(lambda x:x**2start vector transformation... ds_batch_map:
    end vector transformation..."))

================================================================================17:09:10
start scalar transformation...
================================================================================17:09:13
end scalar transformation...
================================================================================17:09:13
start vector transformation...
================================================================================17:09:14
end vector transformation...