tensorflow tfrecord文件存取

Posted 2021-01-13 huwtylv

 

import tensorflow as tf
import numpy as np
import skimage
from skimage import data, io, color

path = "1.tfrecords"

with tf.python_io.TFRecordWriter(path) as writer:
    a = 1024
    b = 10.24
    c = [0.1, 0.2, 0.3]
    d = [[1, 2], [3, 4]]
    e = "Python"
    img = io.imread(‘/data/test/img/a.jpg‘)
    img_shape = img.shape


    c = np.array(c).astype(np.float32).tobytes()
    d = np.array(d).astype(np.int8).tobytes()
    e = bytes(e, encoding=‘utf-8‘)
    img = img.astype(np.int16).tobytes()

    example = tf.train.Example(features=tf.train.Features(feature={
        ‘a‘: tf.train.Feature(int64_list=tf.train.Int64List(value=[a])),
        ‘b‘: tf.train.Feature(float_list=tf.train.FloatList(value=[b])),
        ‘c‘: tf.train.Feature(bytes_list=tf.train.BytesList(value=[c])),
        ‘d‘: tf.train.Feature(bytes_list=tf.train.BytesList(value=[d])),
        ‘e‘: tf.train.Feature(bytes_list=tf.train.BytesList(value=[e])),
        ‘img‘: tf.train.Feature(bytes_list=tf.train.BytesList(value=[img])),

    }))
    writer.write(example.SerializeToString())


# 读取
filename_queue = tf.train.string_input_producer([path])
_, serialized_example = tf.TFRecordReader().read(filename_queue)

features = tf.parse_single_example(serialized_example,
                                   features={
                                       ‘a‘: tf.FixedLenFeature([], tf.int64),
                                       ‘b‘: tf.FixedLenFeature([], tf.float32),
                                       ‘c‘: tf.FixedLenFeature([], tf.string),
                                       ‘d‘: tf.FixedLenFeature([], tf.string),
                                       ‘e‘: tf.FixedLenFeature([], tf.string),
                                       ‘img‘: tf.FixedLenFeature([], tf.string),

                                   })

a = features[‘a‘] # 返回是张量
b = features[‘b‘]

c = features[‘c‘]
c = tf.decode_raw(c, tf.float32)

d = features[‘d‘]
d = tf.decode_raw(d, tf.int8)
d = tf.reshape(d, [2, 2])

e = features[‘e‘]

img = features[‘img‘]
img = tf.decode_raw(img, tf.int16)
img = tf.reshape(img, shape=img_shape)

with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    tf.train.start_queue_runners(sess=sess)
    print(sess.run([a, b, c, d, e]))
    e = sess.run(e)
    print(type(e), bytes.decode(e))
    print(sess.run(img))