def read(data_dir: str,
feature_config: FeatureConfig,
tfrecord_type: str,
file_io: FileIO,
max_sequence_size: int = 0,
batch_size: int = 0,
preprocessing_keys_to_fns: dict = {},
parse_tfrecord: bool = True,
use_part_files: bool = False,
logger: Logger = None,
**kwargs) -> data.TFRecordDataset:
"""
- reads tfrecord data from an input directory
- selects relevant features
- creates X and y data
Args:
data_dir: Path to directory containing csv files to read
feature_config: ml4ir.config.features.Features object extracted from the feature config
tfrecord_type: either example or sequence_example
batch_size: int value specifying the size of the batch
preprocessing_keys_to_fns: dictionary mapping preprocessing keys in the feature_config to functions
parse_tfrecord: whether to parse SequenceExamples into features
logger: logging object
Returns:
tensorflow dataset
"""
parse_fn = get_parse_fn(
feature_config=feature_config,
tfrecord_type=tfrecord_type,
preprocessing_keys_to_fns=preprocessing_keys_to_fns,
max_sequence_size=max_sequence_size,
)
# Get all tfrecord files in directory
tfrecord_files = file_io.get_files_in_directory(
data_dir,
extension="" if use_part_files else ".tfrecord",
prefix="part-" if use_part_files else "",
)
# Parse the protobuf data to create a TFRecordDataset
dataset = data.TFRecordDataset(tfrecord_files)
if parse_tfrecord:
dataset = dataset.map(parse_fn).apply(
data.experimental.ignore_errors())
# Create BatchedDataSet
if batch_size:
dataset = dataset.batch(batch_size, drop_remainder=True)
if logger:
logger.info(
"Created TFRecordDataset from SequenceExample protobufs from {} files : {}"
.format(len(tfrecord_files),
str(tfrecord_files)[:50]))
return dataset
def read(
data_dir: str,
feature_config: FeatureConfig,
tfrecord_type: str,
file_io: FileIO,
max_sequence_size: int = 0,
batch_size: int = 0,
preprocessing_keys_to_fns: dict = {},
parse_tfrecord: bool = True,
use_part_files: bool = False,
logger: Logger = None,
**kwargs
) -> data.TFRecordDataset:
"""
Extract features by reading and parsing TFRecord data
and converting into a TFRecordDataset using the FeatureConfig
Parameters
----------
data_dir : str
path to the directory containing train, validation and test data
feature_config : `FeatureConfig` object
FeatureConfig object that defines the features to be loaded in the dataset
and the preprocessing functions to be applied to each of them
tfrecord_type : {"example", "sequence_example"}
Type of the TFRecord protobuf message to be used for TFRecordDataset
file_io : `FileIO` object
file I/O handler objects for reading and writing data
max_sequence_size : int, optional
maximum number of sequence to be used with a single SequenceExample proto message
The data will be appropriately padded or clipped to fit the max value specified
batch_size : int, optional
size of each data batch
preprocessing_keys_to_fns : dict of (str, function), optional
dictionary of function names mapped to function definitions
that can now be used for preprocessing while loading the
TFRecordDataset to create the RelevanceDataset object
use_part_files : bool, optional
load dataset from part files checked using "part-" prefix
parse_tfrecord : bool, optional
parse the TFRecord string from the dataset;
returns strings as is otherwise
logger : `Logger`, optional
logging handler for status messages
Returns
-------
`TFRecordDataset`
TFRecordDataset loaded from the `data_dir` specified using the FeatureConfig
"""
parse_fn = get_parse_fn(
feature_config=feature_config,
tfrecord_type=tfrecord_type,
preprocessing_keys_to_fns=preprocessing_keys_to_fns,
max_sequence_size=max_sequence_size,
)
# Get all tfrecord files in directory
tfrecord_files = file_io.get_files_in_directory(
data_dir,
extension="" if use_part_files else ".tfrecord",
prefix="part-" if use_part_files else "",
)
# Parse the protobuf data to create a TFRecordDataset
dataset = data.TFRecordDataset(tfrecord_files)
if parse_tfrecord:
# Parallel calls set to AUTOTUNE: improved training performance by 40% with a classification model
dataset = dataset.map(parse_fn, num_parallel_calls=tf.data.experimental.AUTOTUNE).apply(
data.experimental.ignore_errors()
)
# Create BatchedDataSet
if batch_size:
dataset = dataset.batch(batch_size, drop_remainder=True)
if logger:
logger.info(
"Created TFRecordDataset from SequenceExample protobufs from {} files : {}".format(
len(tfrecord_files), str(tfrecord_files)[:50]
)
)
# We apply prefetch as it improved train/test/validation throughput by 30% in some real model training.
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
return dataset
def read(data_dir: str,
feature_config: FeatureConfig,
tfrecord_type: str,
tfrecord_dir: str,
file_io: FileIO,
batch_size: int = 128,
preprocessing_keys_to_fns: dict = {},
use_part_files: bool = False,
max_sequence_size: int = 25,
parse_tfrecord: bool = True,
logger=None,
**kwargs) -> tf.data.TFRecordDataset:
"""
Create a TFRecordDataset from directory of CSV files using the FeatureConfig
Current execution plan:
1. Load CSVs as pandas dataframes
2. Convert each query into tf.train.SequenceExample protobufs
3. Write the protobufs into a .tfrecord file
4. Load .tfrecord file into a TFRecordDataset and parse the protobufs
Parameters
data_dir : str
Path to directory containing csv files to read
feature_config : FeatureConfig object
FeatureConfig object that defines the features to be loaded in the dataset
and the preprocessing functions to be applied to each of them
tfrecord_dir : str
Path to directory where the serialized .tfrecord files will be stored
batch_size : int
value specifying the size of the data batch
use_part_files : bool
load dataset from part files checked using "part-" prefix
max_sequence_size : int
value specifying max number of records per query
logger : Logger object
logging handler to print and save status messages
Returns
-------
`TFRecordDataset` object
tensorflow TFRecordDataset loaded from the CSV file
"""
csv_files: List[str] = file_io.get_files_in_directory(
data_dir,
extension="" if use_part_files else ".csv",
prefix="part-" if use_part_files else "",
)
# Create a directory for storing tfrecord files
file_io.make_directory(tfrecord_dir, clear_dir=True)
# Write tfrecord files
tfrecord_writer.write_from_files(
csv_files=csv_files,
tfrecord_file=os.path.join(tfrecord_dir, TFRECORD_FILE),
feature_config=feature_config,
tfrecord_type=tfrecord_type,
file_io=file_io,
logger=logger,
)
dataset = tfrecord_reader.read(
data_dir=tfrecord_dir,
feature_config=feature_config,
tfrecord_type=tfrecord_type,
max_sequence_size=max_sequence_size,
batch_size=batch_size,
preprocessing_keys_to_fns=preprocessing_keys_to_fns,
parse_tfrecord=parse_tfrecord,
file_io=file_io,
logger=logger,
)
return dataset
def read(data_dir: str,
feature_config: FeatureConfig,
tfrecord_type: str,
tfrecord_dir: str,
file_io: FileIO,
batch_size: int = 128,
preprocessing_keys_to_fns: dict = {},
use_part_files: bool = False,
max_sequence_size: int = 25,
parse_tfrecord: bool = True,
logger=None,
keep_additional_info=0,
non_zero_features_only=1,
**kwargs) -> tf.data.TFRecordDataset:
"""
- reads ranklib-formatted data from an input directory
- selects relevant features
- creates Dataset X and y
Current execution plan:
1. Convert ranklib to a dataframe
2. Convert each query into tf.train.SequenceExample protobufs
3. Write the protobufs into a .tfrecord file
4. Load .tfrecord file into a TFRecordDataset and parse the protobufs
Parameters
----------
data_dir: str
Path to directory containing csv files to read
feature_config: ml4ir.config.features.FeatureConfig object
FeatureConfig object extracted from the feature config
tfrecord_dir: str
Path to directory where the serialized .tfrecord files will be stored
batch_size: int
Value specifying the size of the batch
use_part_files: bool
Value specifying whether to look for part files
max_sequence_size: int
Value specifying max number of records per query
logger: logging object
logging object
keep_additional_info: int
Option to keep additional info (All info after the "#") 1 to keep, 0 to ignore
non_zero_features_only: int
Only non zero features are stored. 1 for yes, 0 otherwise
Returns
-------
tensorflow TFRecordDataset
Processed dataset
"""
ranklib_files: List[str] = file_io.get_files_in_directory(
data_dir,
extension="" if use_part_files else ".txt",
prefix="part-" if use_part_files else "",
)
gl_2_clicks = False
# Create a directory for storing tfrecord files
file_io.make_directory(tfrecord_dir, clear_dir=True)
#Convert input ranklib file to dataframe
df = pd.concat([
ranklib_helper.convert(f, keep_additional_info, gl_2_clicks,
non_zero_features_only,
feature_config.get_query_key()['name'],
feature_config.get_label()['name'])
for f in ranklib_files
])
#Write tfrecord files
tfrecord_writer.write_from_df(df=df,
tfrecord_file=os.path.join(
tfrecord_dir, TFRECORD_FILE),
feature_config=feature_config,
tfrecord_type=tfrecord_type,
logger=logger)
dataset = tfrecord_reader.read(
data_dir=tfrecord_dir,
feature_config=feature_config,
tfrecord_type=tfrecord_type,
max_sequence_size=max_sequence_size,
batch_size=batch_size,
preprocessing_keys_to_fns=preprocessing_keys_to_fns,
parse_tfrecord=parse_tfrecord,
file_io=file_io,
logger=logger,
)
return dataset
def read(data_dir: str,
feature_config: FeatureConfig,
tfrecord_type: str,
tfrecord_dir: str,
file_io: FileIO,
batch_size: int = 128,
preprocessing_keys_to_fns: dict = {},
use_part_files: bool = False,
max_sequence_size: int = 25,
parse_tfrecord: bool = True,
logger=None,
**kwargs) -> tf.data.TFRecordDataset:
"""
- reads csv-formatted data from an input directory
- selects relevant features
- creates Dataset X and y
Current execution plan:
1. Load CSVs as pandas dataframes
2. Convert each query into tf.train.SequenceExample protobufs
3. Write the protobufs into a .tfrecord file
4. Load .tfrecord file into a TFRecordDataset and parse the protobufs
Args:
- data_dir: Path to directory containing csv files to read
- feature_config: ml4ir.config.features.FeatureConfig object extracted from the feature config
- tfrecord_dir: Path to directory where the serialized .tfrecord files will be stored
- batch_size: int value specifying the size of the batch
- use_part_files: bool value specifying whether to look for part files
- max_sequence_size: int value specifying max number of records per query
- logger: logging object
Returns:
tensorflow TFRecordDataset
"""
csv_files: List[str] = file_io.get_files_in_directory(
data_dir,
extension="" if use_part_files else ".csv",
prefix="part-" if use_part_files else "",
)
# Create a directory for storing tfrecord files
file_io.make_directory(tfrecord_dir, clear_dir=True)
# Write tfrecord files
tfrecord_writer.write_from_files(
csv_files=csv_files,
tfrecord_file=os.path.join(tfrecord_dir, TFRECORD_FILE),
feature_config=feature_config,
tfrecord_type=tfrecord_type,
file_io=file_io,
logger=logger,
)
dataset = tfrecord_reader.read(
data_dir=tfrecord_dir,
feature_config=feature_config,
tfrecord_type=tfrecord_type,
max_sequence_size=max_sequence_size,
batch_size=batch_size,
preprocessing_keys_to_fns=preprocessing_keys_to_fns,
parse_tfrecord=parse_tfrecord,
file_io=file_io,
logger=logger,
)
return dataset
