tf.io.TFRecordWriter

A class to write records to a TFRecords file.

Used in the notebooks

Used in the tutorials

TFRecords tutorial

TFRecords is a binary format which is optimized for high throughput data retrieval, generally in conjunction with tf.data. TFRecordWriter is used to write serialized examples to a file for later consumption. The key steps are:

Ahead of time:

A minimal example is given below:

import tempfile
example_path = os.path.join(tempfile.gettempdir(), "example.tfrecords")
np.random.seed(0)
# Write the records to a file.
with tf.io.TFRecordWriter(example_path) as file_writer:
  for _ in range(4):
    x, y = np.random.random(), np.random.random()

    record_bytes = tf.train.Example(features=tf.train.Features(feature={
        "x": tf.train.Feature(float_list=tf.train.FloatList(value=[x])),
        "y": tf.train.Feature(float_list=tf.train.FloatList(value=[y])),
    })).SerializeToString()
    file_writer.write(record_bytes)
# Read the data back out.
def decode_fn(record_bytes):
  return tf.io.parse_single_example(
      # Data
      record_bytes,

      # Schema
      {"x": tf.io.FixedLenFeature([], dtype=tf.float32),
       "y": tf.io.FixedLenFeature([], dtype=tf.float32)}
  )
for batch in tf.data.TFRecordDataset([example_path]).map(decode_fn):
  print("x = {x:.4f},  y = {y:.4f}".format(**batch))
x = 0.5488,  y = 0.7152
x = 0.6028,  y = 0.5449
x = 0.4237,  y = 0.6459
x = 0.4376,  y = 0.8918

This class implements __enter__ and __exit__, and can be used in with blocks like a normal file. (See the usage example above.)

path The path to the TFRecords file.
options (optional) String specifying compression type, TFRecordCompressionType, or TFRecordOptions object.

IOError If path cannot be opened for writing.
ValueError If valid compression_type can't be determined from options.

Methods

close

View source

Close the file.

flush

View source

Flush the file.

write

View source

Write a string record to the file.

Args
record str

__enter__

enter(self: object) -> object

__exit__

exit(self: tensorflow.python._pywrap_record_io.RecordWriter, *args) -> None