ParquetDataset

Description

ParquetDataset supports reading data from parquet files.
ParquetDataset supports reading parquet files from local or S3/OSS/HDFS file systems.

UserAPI

Environment Variable

# If ARROW_NUM_THREADS > 0, specified number of threads will be used.
# If ARROW_NUM_THREADS = 0, no threads will be used.
# If ARROW_NUM_THREADS < 0, all threads will be used.
os.environ['ARROW_NUM_THREADS'] = '2'

ParquetDataset API

class ParquetDataset(dataset_ops.DatasetV2):
  def __init__(
      self, filenames,
      batch_size=1,
      fields=None,
      partition_count=1,
      partition_index=0,
      drop_remainder=False,
      num_parallel_reads=None,
      num_sequential_reads=1):

# Create a `ParquetDataset` from filenames dataset.
def read_parquet(
    batch_size,
    fields=None,
    partition_count=1,
    partition_index=0,
    drop_remainder=False,
    num_parallel_reads=None,
    num_sequential_reads=1):

filenames: the filename of parquet file, This parameter can receive the following types.
- A 0-D or 1-D tf.string tensor
- string
- list or tuple of string
- Dataset containing one or more filenames.
batch_size: (Optional.) Maxium number of samples in an output batch.

fields: (Optional.) List of DataFrame fields.

`filenames` parameter type	`fields` parameter requirement	`fields` parameter type
`Tensor`/`DataSet`	required	`DataFrame.Field`/`list` or `tuple` of `DataFrame.Field`
`string`/`list` or `tuple` of `string`	optional, the default value means read all columns	`DataFrame.Field`/`list` or `tuple` of `DataFrame.Field`/`string`/`list` or `tuple` of `string`

partition_count: (Optional.) Count of row group partitions.
partition_index: (Optional.) Index of row group partitions.
drop_remainder: (Optional.) If True, only keep batches with exactly batch_size samples.
num_parallel_reads: (Optional.) A tf.int64 scalar representing the number of files to read in parallel. Defaults to reading files sequentially.
num_sequential_reads: (Optional.) A tf.int64 scalar representing the number of batches to read in sequential. Defaults to 1.

DataFrame

A data frame is a table consisting of multiple named columns. A named column has a logical data type and a physical data type.

Logical Type of DataFrame

Logical Type	Output Type
Scalar	`tf.Tensor`/`DataFrame.Value`
Fixed-Length List	`tf.Tensor`/`DataFrame.Value`
Variable-Length List	`tf.SparseTensor`/`DataFrame.Value`
Variable-Length Nested List	`tf.SparseTensor`/`DataFrame.Value`

Physical Type of DataFrame

Category	Physical Type
Integers	`int64` `uint64` `int32` `uint32` `int8` `uint8`
Numerics	`float64` `float32` `float16`
Text	`string`

DataFrame API

class DataFrame(object):
    class Field(object):
        def __init__(self, name,
            type=None,
            ragged_rank=None,
            shape=None):

    class Value(collections.namedtuple(
        'DataFrameValue', ['values', 'nested_row_splits'])):
        def to_sparse(self, name=None):

# Convert values to tensors or sparse tensors from input dataset.
def to_sparse(num_parallel_calls=None):

DataFrame.Field API

name: Name of column.
type: data type of column, such as tf.int64
ragged_rank: (optional.) Specify the number of nesting levels when column is a nested list
shape: (optional.) Specify the shape of column when column is a fixed-length list

Attention: For fix-length list, only the shape needs to be specified.

DataFrame.Value Conversion API (Use according to the actual situation)

Since there may be DataFrame.Value types in the output of ParquetDataset that cannot be accessed by model directly, it needs to convert the DataFrame.Value to SparseTensor. Please use the to_sparse API for conversion.

import tensorflow as tf
from tensorflow.python.data.experimental.ops import parquet_dataset_ops
from tensorflow.python.data.experimental.ops import dataframe

ds = parquet_dataset_ops.ParquetDataset(...)
ds.apply(dataframe.to_sparse())
...

Examples

1. Read from one file on local filesystem

import tensorflow as tf
from tensorflow.python.data.experimental.ops import parquet_dataset_ops

# Read from a parquet file.
ds = parquet_dataset_ops.ParquetDataset('/path/to/f1.parquet',
                                        batch_size=1024)
ds = ds.prefetch(4)
it = tf.data.make_one_shot_iterator(ds)
batch = it.get_next()
# {'a': tensora, 'c': tensorc}

2. Read from filenames dataset

import tensorflow as tf
from tensorflow.python.data.experimental.ops import parquet_dataset_ops

filenames = tf.data.Dataset.from_generator(func, tf.string, tf.TensorShape([]))
# Define data frame fields.
fields = [
    parquet_dataset_ops.DataFrame.Field('A', tf.int64),
    parquet_dataset_ops.DataFrame.Field('C', tf.int64, ragged_rank=1)]
# Read from parquet files by reading upstream filename dataset.
ds = filenames.apply(parquet_dataset_ops.read_parquet(1024, fields=fields))
ds = ds.prefetch(4)
it = tf.data.make_one_shot_iterator(ds)
batch = it.get_next()
# {'a': tensora, 'c': tensorc}

3. Read from files on HDFS

import tensorflow as tf
from tensorflow.python.data.experimental.ops import parquet_dataset_ops

# Read from parquet files on remote services for selected fields.
ds = parquet_dataset_ops.ParquetDataset(
    ['hdfs://host:port/path/to/f3.parquet'],
    batch_size=1024,
    fields=['a', 'c'])
ds = ds.prefetch(4)
it = tf.data.make_one_shot_iterator(ds)
batch = it.get_next()
# {'a': tensora, 'c': tensorc}