tensorflow 变长序列存储实例

问题

问题是这样的，要把一个数组存到tfrecord中，然后读取

a = np.array([[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196],
[0, 38, 79, 157],
[0, 49, 89, 147, 177],
[0, 32, 73, 145]])

图片我都存储了，这个不还是小意思，一顿操作

import tensorflow as tf
import numpy as np
def _int64_feature(value):
if not isinstance(value,list):
value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

# Write an array to TFrecord.
# a is an array which contains lists of variant length.a = np.array([[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196],
[0, 38, 79, 157],
[0, 49, 89, 147, 177],
[0, 32, 73, 145]])

writer = tf.python_io.TFRecordWriter('file')

for i in range(a.shape[0]):
feature = {'i' : _int64_feature(i),
'data': _int64_feature(a[i])}

# Create an example protocol buffer
example = tf.train.Example(features=tf.train.Features(feature=feature))

# Serialize to string and write on the file
writer.write(example.SerializeToString())

writer.close()

# Use Dataset API to read the TFRecord file.
filenames = ["file"]
dataset = tf.data.TFRecordDataset(filenames)
def _parse_function(example_proto):
keys_to_features = {'i':tf.FixedLenFeature([],tf.int64),
'data':tf.FixedLenFeature([],tf.int64)}
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
return parsed_features['i'], parsed_features['data']

dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(buffer_size=1)
dataset = dataset.repeat()
dataset = dataset.batch(1)
iterator = dataset.make_one_shot_iterator()
i, data = iterator.get_next()
with tf.Session() as sess:
print(sess.run([i, data]))
print(sess.run([i, data]))
print(sess.run([i, data]))

报了奇怪的错误，Name: <unknown>, Key: data, Index: 0. Number of int64 values != expected. Values size: 6 but output shape: [] 这意思是我数据长度为6，但是读出来的是[]，这到底是哪里错了，我先把读取的代码注释掉，看看tfreocrd有没有写成功，发现写成功了，这就表明是读取的问题，我怀疑是因为每次写入的长度是变化的原因，但是又有觉得不是，因为图片的尺寸都是不同的，我还是可以读取的，百思不得其解的时候我发现存储图片的时候是img.tobytes(),我把一个数组转换成了bytes，而且用的也是bytes存储，是不是tensorflow会把这个bytes当成一个元素，虽然每个图片的size不同，但是tobytes后tensorflow都会当成一个元素，然后读取的时候再根据(height,width,channel)来解析成图片。

我来试试不存为int64，而是存为bytes。 又是一顿厉害的操作

数据转为bytes

# -*- coding: utf-8 -*-

import tensorflow as tf
import numpy as np

def _byte_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
def _int64_feature(value):
if not isinstance(value,list):
value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))
# Write an array to TFrecord.
# a is an array which contains lists of variant length.a = np.array([[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196],
[0, 38, 79, 157],
[0, 49, 89, 147, 177],
[0, 32, 73, 145]])

writer = tf.python_io.TFRecordWriter('file')

for i in range(a.shape[0]): # i = 0 ~ 4
feature = {'len' : _int64_feature(len(a[i])), # 将无意义的i改成len，为了后面还原
'data': _byte_feature(np.array(a[i]).tobytes())} # 我也不知道为什么a[i]是list（后面就知道了），要存bytes需要numpy一下

# Create an example protocol buffer
example = tf.train.Example(features=tf.train.Features(feature=feature))

# Serialize to string and write on the file
writer.write(example.SerializeToString())

writer.close()

#
# Use Dataset API to read the TFRecord file.
filenames = ["file"]
dataset = tf.data.TFRecordDataset(filenames)
def _parse_function(example_proto):
keys_to_features = {'len':tf.FixedLenFeature([],tf.int64),
'data':tf.FixedLenFeature([],tf.string)} # 改成string
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
return parsed_features['len'], parsed_features['data']

dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(buffer_size=1)
dataset = dataset.repeat()
dataset = dataset.batch(1)
iterator = dataset.make_one_shot_iterator()
i, data = iterator.get_next()
with tf.Session() as sess:
print(sess.run([i, data]))
print(sess.run([i, data]))
print(sess.run([i, data]))

"""
[array([6], dtype=int64), array([b'\x00\x00\x00\x006\x00\x00\x00[\x00\x00\x00\x99\x00\x00\x00\xb1\x00\x00\x00\x01\x00\x00\x00'],
dtype=object)]
[array([5], dtype=int64), array([b'\x00\x00\x00\x002\x00\x00\x00Y\x00\x00\x00\x93\x00\x00\x00\xc4\x00\x00\x00'],
dtype=object)]
[array([4], dtype=int64), array([b'\x00\x00\x00\x00&\x00\x00\x00O\x00\x00\x00\x9d\x00\x00\x00'],
dtype=object)]
"""

bytes数据解码

如愿的输出来了，但是这个bytes我该如何解码呢

方法一，我们自己解析

a,b= sess.run([i,data])
c = np.frombuffer(b[0],dtype=np.int,count=a[0])

方法二使用tensorflow的解析函数

def _parse_function(example_proto):
keys_to_features = {'len':tf.FixedLenFeature([],tf.int64),
'data':tf.FixedLenFeature([],tf.string)} # 改成string
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
dat = tf.decode_raw(parsed_features['data'],tf.int64) # 用的是这个解析函数，我们使用int64的格式存储的，解析的时候也是转换为int64
return parsed_features['len'], dat
"""
[array([6]), array([[ 0, 54, 91, 153, 177, 1]])]
[array([5]), array([[ 0, 50, 89, 147, 196]])]
[array([4]), array([[ 0, 38, 79, 157]])]
"""

可以看到是二维数组，这是因为我们使用的是batch输出，虽然我们的bathc_size=1，但是还是会以二维list的格式输出。我手贱再来修改点东西，

def _parse_function(example_proto):
keys_to_features = {'len':tf.FixedLenFeature([1],tf.int64),
'data':tf.FixedLenFeature([1],tf.string)}
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
dat = tf.decode_raw(parsed_features['data'],tf.int64)
return parsed_features['len'], dat

"""
[array([[6]]), array([[[ 0, 54, 91, 153, 177, 1]]])]
[array([[5]]), array([[[ 0, 50, 89, 147, 196]]])]
[array([[4]]), array([[[ 0, 38, 79, 157]]])]
"""

呦呵，又变成3维的了，让他报个错试试

def _parse_function(example_proto):
keys_to_features = {'len':tf.FixedLenFeature([2],tf.int64), # 1 修改为 2
'data':tf.FixedLenFeature([1],tf.string)} # 改成string
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
return parsed_features['len'], parsed_features['data']

"""
InvalidArgumentError: Key: len. Can't parse serialized Example.
[[Node: ParseSingleExample/ParseSingleExample = ParseSingleExample[Tdense=[DT_STRING, DT_INT64], dense_keys=["data", "len"], dense_shapes=[[1], [2]], num_sparse=0, sparse_keys=[], sparse_types=[]](arg0, ParseSingleExample/Const, ParseSingleExample/Const_1)]]
[[Node: IteratorGetNext_22 = IteratorGetNext[output_shapes=[[?,2], [?,1]], output_types=[DT_INT64, DT_STRING], _device="/job:localhost/replica:0/task:0/device:CPU:0"](OneShotIterator_22)]]
"""

可以看到dense_keys=["data", "len"], dense_shapes=[[1], [2]],，tf.FixedLenFeature是读取固定长度的数据，我猜测[]的意思就是读取全部数据，[1]就是读取一个数据，每个数据可能包含多个数据，形如[[1，2],[3，3，4],[2]....]，哈哈这都是我瞎猜的，做我女朋友好不好。

tensorflow 变长数组存储

反正是可以读取了。但是如果是自己定义的变长数组，每次都要自己解析，这样很麻烦（我瞎遍的），所以tensorflow就定义了变长数组的解析方法tf.VarLenFeature，我们就不需要把边长数组变为bytes再解析了，又是一顿操作

import tensorflow as tf
import numpy as np
def _int64_feature(value):
if not isinstance(value,list):
value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

# Write an array to TFrecord.
# a is an array which contains lists of variant length.a = np.array([[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196],
[0, 38, 79, 157],
[0, 49, 89, 147, 177],
[0, 32, 73, 145]])

writer = tf.python_io.TFRecordWriter('file')

for i in range(a.shape[0]): # i = 0 ~ 4
feature = {'i' : _int64_feature(i),
'data': _int64_feature(a[i])}

# Create an example protocol buffer
example = tf.train.Example(features=tf.train.Features(feature=feature))

# Serialize to string and write on the file
writer.write(example.SerializeToString())

writer.close()

# Use Dataset API to read the TFRecord file.
filenames = ["file"]
dataset = tf.data.TFRecordDataset(filenames)
def _parse_function(example_proto):
keys_to_features = {'i':tf.FixedLenFeature([],tf.int64),
'data':tf.VarLenFeature(tf.int64)}
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
return parsed_features['i'], tf.sparse_tensor_to_dense(parsed_features['data'])

dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(buffer_size=1)
dataset = dataset.repeat()
dataset = dataset.batch(1)
iterator = dataset.make_one_shot_iterator()
i, data = iterator.get_next()
with tf.Session() as sess:
print(sess.run([i, data]))
print(sess.run([i, data]))
print(sess.run([i, data]))

"""
[array([0], dtype=int64), array([[ 0, 54, 91, 153, 177, 1]], dtype=int64)]
[array([1], dtype=int64), array([[ 0, 50, 89, 147, 196]], dtype=int64)]
[array([2], dtype=int64), array([[ 0, 38, 79, 157]], dtype=int64)]
"""

batch输出

输出还是数组，哈哈哈。再来一波操作

dataset = dataset.batch(2)
"""
Cannot batch tensors with different shapes in component 1. First element had shape [6] and element 1 had shape [5].
"""

这是因为一个batch中数据的shape必须是一致的，第一个元素长度为6，第二个元素长度为5，就会报错。办法就是补成一样的长度，在这之前先测试点别的

a = np.array([[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196],
[0, 38, 79, 157],
[0, 49, 89, 147, 177],
[0, 32, 73, 145]])

for i in range(a.shape[0]):
print(type(a[i]))

"""
<class 'list'>
<class 'list'>
<class 'list'>
<class 'list'>
<class 'list'>
"""

可以发现长度不一的array每一个数据是list（一开始我以为是object）。然后补齐

a = np.array([[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196,0],
[0, 38, 79, 157,0,0],
[0, 49, 89, 147, 177,0],
[0, 32, 73, 145,0,0]])

for i in range(a.shape[0]):
print(type(a[i]))

"""
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>
<class 'numpy.ndarray'>
"""

返回的是numpy。为什么要做这件事呢？

def _int64_feature(value):
if not isinstance(value,list):
value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

tensorflow要求我们输入的是list或者直接是numpy.ndarry，如果是list中包含numpy.ndarry [numpy.ndarry]就会报错。上面的那个数组时边长的，返回的时list，没有什么错误，我们补齐看看

a = np.array([[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196,0],
[0, 38, 79, 157,0,0],
[0, 49, 89, 147, 177,0],
[0, 32, 73, 145,0,0]])

"""
TypeError: only size-1 arrays can be converted to Python scalars
"""

这就是因为返回的不是list，而是numpy.ndarry,而_int64_feature函数中先判断numpy.ndarry不是list，所以转成了[numpy.ndarry]就报错了。可以做些修改，一种方法是将numpy.ndarry转为list

for i in range(a.shape[0]): # i = 0 ~ 4
feature = {'i' : _int64_feature(i),
'data': _int64_feature(a[i].tolist())}

这样补齐了我们就可以修改batch的值了

dataset = dataset.batch(2)

"""
[array([0, 2], dtype=int64), array([[ 0, 54, 91, 153, 177, 1],
[ 0, 38, 79, 157, 0, 0]], dtype=int64)]
[array([1, 3], dtype=int64), array([[ 0, 50, 89, 147, 196, 0],
[ 0, 49, 89, 147, 177, 0]], dtype=int64)]
[array([4, 0], dtype=int64), array([[ 0, 32, 73, 145, 0, 0],
[ 0, 54, 91, 153, 177, 1]], dtype=int64)]
"""

当然tensorflow不会让我自己补齐，已经提供了补齐函数padded_batch，

# -*- coding: utf-8 -*-

import tensorflow as tf
def _int64_feature(value):
if not isinstance(value,list):
value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

a = [[0, 54, 91, 153, 177,1],
[0, 50, 89, 147, 196],
[0, 38, 79, 157],
[0, 49, 89, 147, 177],
[0, 32, 73, 145]]

writer = tf.python_io.TFRecordWriter('file')

for v in a: # i = 0 ~ 4
feature = {'data': _int64_feature(v)}

# Create an example protocol buffer
example = tf.train.Example(features=tf.train.Features(feature=feature))

# Serialize to string and write on the file
writer.write(example.SerializeToString())

writer.close()

# Use Dataset API to read the TFRecord file.
filenames = ["file"]
dataset = tf.data.TFRecordDataset(filenames)
def _parse_function(example_proto):
keys_to_features = {'data':tf.VarLenFeature(tf.int64)}
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
return tf.sparse_tensor_to_dense( parsed_features['data'])

dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(buffer_size=1)
dataset = dataset.repeat()
dataset = dataset.padded_batch(2,padded_shapes=([None]))
iterator = dataset.make_one_shot_iterator()
data = iterator.get_next()
with tf.Session() as sess:
print(sess.run([data]))
print(sess.run([data]))
print(sess.run([data]))

"""
[array([[ 0, 54, 91, 153, 177, 1],
[ 0, 50, 89, 147, 196, 0]])]
[array([[ 0, 38, 79, 157, 0],
[ 0, 49, 89, 147, 177]])]
[array([[ 0, 32, 73, 145, 0, 0],
[ 0, 54, 91, 153, 177, 1]])]
"""

可以看到的确是自动补齐了。

图片batch

直接来测试一下图片数据

# -*- coding: utf-8 -*-

import tensorflow as tf
import matplotlib.pyplot as plt
def _byte_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

files = tf.gfile.Glob('*.jpeg')
writer = tf.python_io.TFRecordWriter('file')
for file in files:

with tf.gfile.FastGFile(file,'rb') as f:
img_buff = f.read()
feature = {'img': _byte_feature(tf.compat.as_bytes(img_buff))}
example = tf.train.Example(features=tf.train.Features(feature=feature))
writer.write(example.SerializeToString())
writer.close()

filenames = ["file"]
dataset = tf.data.TFRecordDataset(filenames)
def _parse_function(example_proto):
keys_to_features = {'img':tf.FixedLenFeature([], tf.string)}
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
image = tf.image.decode_jpeg(parsed_features['img'])
return image

dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(buffer_size=1)
dataset = dataset.repeat()
dataset = dataset.batch(2)
iterator = dataset.make_one_shot_iterator()
image = iterator.get_next()

with tf.Session() as sess:
img = sess.run([image])
print(len(img))
print(img[0].shape)
plt.imshow(img[0][0])

"""
Cannot batch tensors with different shapes in component 0. First element had shape [440,440,3] and element 1 had shape [415,438,3].
"""

看到了没有，一个batch中图片的尺寸不同，就不可以batch了，我们必须要将一个batch的图片resize成相同的代大小。

def _parse_function(example_proto):
keys_to_features = {'img':tf.FixedLenFeature([], tf.string)}
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
image = tf.image.decode_jpeg(parsed_features['img'])
image = tf.image.convert_image_dtype(image,tf.float32)# 直接resize，会将uint8转为float类型，但是plt.imshow只能显示uint8或者0-1之间float类型，这个函数就是将uint8转为0-1之间的float类型，相当于除以255.0
image = tf.image.resize_images(image,(224,224))
return image

但是有时候我们希望输入图片尺寸是不一样的，不需要reize，这样只能将batch_size=1。一个batch中的图片shape必须是一样的，我们可以这样折中训练，使用tensorflow提供的动态填充接口，将一个batch中的图片填充为相同的shape。

dataset = dataset.padded_batch(2,padded_shapes=([None,None,3]))

如果我们想要将图片的名称作为标签保存下来要怎么做呢？

# -*- coding: utf-8 -*-

import tensorflow as tf
import matplotlib.pyplot as plt
import os

out_charset="ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789"

def _byte_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

def _int64_feature(values):
if not isinstance(values,list):
values = [values]
return tf.train.Feature(int64_list=tf.train.Int64List(value=values))

files = tf.gfile.Glob('*.jpg')
writer = tf.python_io.TFRecordWriter('file')
for file in files:
with tf.gfile.FastGFile(file,'rb') as f:
img_buff = f.read()
filename = os.path.basename(file).split('.')[0]
label = list(map(lambda x:out_charset.index(x),filename))
feature = {'label':_int64_feature(label),
'filename':_byte_feature(tf.compat.as_bytes(filename)),
'img': _byte_feature(tf.compat.as_bytes(img_buff))}
example = tf.train.Example(features=tf.train.Features(feature=feature))
writer.write(example.SerializeToString())
writer.close()

filenames = ["file"]
dataset = tf.data.TFRecordDataset(filenames)
def _parse_function(example_proto):
keys_to_features = {
'label':tf.VarLenFeature(tf.int64),
'filename':tf.FixedLenFeature([],tf.string),
'img':tf.FixedLenFeature([], tf.string)}
parsed_features = tf.parse_single_example(example_proto, keys_to_features)
label = tf.sparse_tensor_to_dense(parsed_features['label'])
filename = parsed_features['filename']
image = tf.image.decode_jpeg(parsed_features['img'])
return image,label,filename

dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(buffer_size=1)
dataset = dataset.repeat()
dataset = dataset.padded_batch(3,padded_shapes=([None,None,3],[None],[]))
#因为返回有三个，所以每一个都要有padded_shapes,但是解码后的image和label都是变长的
#所以需要pad None,而filename没有解码，返回来是byte类型的，只有一个值，所以不需要pad
iterator = dataset.make_one_shot_iterator()
image,label,filename = iterator.get_next()

with tf.Session() as sess:
print(label.eval())

瞎试

如果写入的数据是一个list会是怎样呢

a = np.arange(16).reshape(2,4,2)

"""
TypeError: [0, 1] has type list, but expected one of: int, long
"""

不过想想也是，tf.train.Feature(int64_list=tf.train.Int64List(value=value))这个函数就是存储数据类型为int64的list的。但是如果我们要存储词向量该怎么办呢？例如一句话是一个样本s1='我爱你',假如使用one-hot编码，我=[0,0,1],爱=[0,1,0],你=[1,0,0],s1=[[0,0,1],[0,1,0],[1,0,0]]。这一个样本该怎么存储呢？

以上这篇tensorflow 变长序列存储实例就是小编分享给大家的全部内容了，希望能给大家一个参考

您可能感兴趣的文章:

• Tensorflow实现部分参数梯度更新操作
• TensorFlow损失函数专题详解
• tensorflow如何继续训练之前保存的模型实例
• tensorflow模型继续训练 fineturn实例
• Tensorflow的梯度异步更新示例