Read individual records from different formats.
Readers iterate records from a dataset in storage and load the data into DataVec. The usefulness of readers beyond individual entries in a dataset includes: what if you wanted to train a text generator on a corpus? Or programmatically compose two entries together to form a new record? Reader implementations are useful for complex file types or distributed storage mechanisms.
Readers return Writable classes that describe each column in a Record. These classes are used to convert each record to a tensor/ND-Array format.
Each reader implementation extends BaseRecordReader and provides a simple API for selecting the next record in a dataset, acting similarly to iterators.
Useful methods include:
next: Return a batch of Writable.
nextRecord: Return a single Record, optionally with RecordMetaData.
You can hook a custom RecordListener to a record reader for debugging or visualization purposes. Pass your custom listener to the addListener base method immediately after initializing your class.
Types of readers
ComposableRecordReader
RecordReader for each pipeline. Individual record is a concatenation of the two collections. Create a recordreader that takes recordreaders and iterates over them and concatenates them hasNext would be the & of all the recordreaders concatenation would be next & addAll on the collection return one record
initialize
ConcatenatingRecordReader
Combine multiple readers into a single reader. Records are read sequentially - thus if the first reader has 100 records, and the second reader has 200 records, ConcatenatingRecordReader will have 300 records.
FileRecordReader
File reader/writer
getCurrentLabel
Return the current label. The index of the current file’s parent directory in the label list
return The index of the current file’s parent directory
LineRecordReader
Reads files line by line
CollectionRecordReader
Collection record reader. Mainly used for testing.
CollectionSequenceRecordReader
Collection record reader for sequences. Mainly used for testing.
initialize
param records Collection of sequences. For example, List<List<List>> where the inner two lists are a sequence, and the outer list/collection is a list of sequences
ListStringRecordReader
Iterates through a list of strings return a record.
initialize
Called once at initialization.
param split the split that defines the range of records to read
throws InterruptedException
initialize
Called once at initialization.
param conf a configuration for initialization
param split the split that defines the range of records to read
hasNext
Get the next record
return The list of next record
reset
List of label strings
nextRecord
Load the record from the given DataInputStream Unlike {- link #next()} the internal state of the RecordReader is not modified Implementations of this method should not close the DataInputStream
throws IOException if error occurs during reading from the input stream
close
Closes this stream and releases any system resources associated with it. If the stream is already closed then invoking this method has no effect.
As noted in {- link AutoCloseable#close()}, cases where the close may fail require careful attention. It is strongly advised to relinquish the underlying resources and to internally mark the {- code Closeable} as closed, prior to throwing the {- code IOException}.
throws IOException if an I/O error occurs
setConf
Set the configuration to be used by this object.
getConf
Return the configuration used by this object.
CSVRecordReader
Simple csv record reader.
initialize
Skip first n lines
param skipNumLines the number of lines to skip
CSVRegexRecordReader
A CSVRecordReader that can split each column into additional columns using regexs.
CSVSequenceRecordReader
CSV Sequence Record Reader This reader is intended to read sequences of data in CSV format, where each sequence is defined in its own file (and there are multiple files) Each line in the file represents one time step
CSVVariableSlidingWindowRecordReader
A sliding window of variable size across an entire CSV.
In practice the sliding window size starts at 1, then linearly increase to maxLinesPer sequence, then linearly decrease back to 1.
initialize
No-arg constructor with the default number of lines per sequence (10)
LibSvmRecordReader
Record reader for libsvm format, which is closely related to SVMLight format. Similar to scikit-learn we use a single reader for both formats, so this class is a subclass of SVMLightRecordReader.
Further details on the format can be found at
MatlabRecordReader
Matlab record reader
SVMLightRecordReader
Record reader for SVMLight format, which can generally be described as
LABEL INDEX:VALUE INDEX:VALUE …
SVMLight format is well-suited to sparse data (e.g., bag-of-words) because it omits all features with value zero.
We support an “extended” version that allows for multiple targets (or labels) separated by a comma, as follows:
LABEL1,LABEL2,… INDEX:VALUE INDEX:VALUE …
This can be used to represent either multitask problems or multilabel problems with sparse binary labels (controlled via the “MULTILABEL” configuration option).
Like scikit-learn, we support both zero-based and one-based indexing.
Further details on the format can be found at
initialize
Must be called before attempting to read records.
param conf DataVec configuration
setConf
Set configuration.
param conf DataVec configuration
throws InterruptedException
hasNext
Helper function to help detect lines that are commented out. May read ahead and cache a line.
nextRecord
Return next record as list of Writables.
RegexLineRecordReader
RegexLineRecordReader: Read a file, one line at a time, and split it into fields using a regex. To load an entire file using a
Example: Data in format “2016-01-01 23:59:59.001 1 DEBUG First entry message!”
using regex String “(\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}.\d{3}) (\d+) ([A-Z]+) (.)”
would be split into 4 Text writables: [“2016-01-01 23:59:59.001”, “1”, “DEBUG”, “First entry message!”]
RegexSequenceRecordReader
RegexSequenceRecordReader: Read an entire file (as a sequence), one line at a time and split each line into fields using a regex.
Example: Data in format “2016-01-01 23:59:59.001 1 DEBUG First entry message!”
using regex String “(\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}.\d{3}) (\d+) ([A-Z]+) (.)”
would be split into 4 Text writables: [“2016-01-01 23:59:59.001”, “1”, “DEBUG”, “First entry message!”]
lines that don’t match the provided regex can result in an exception (FailOnInvalid), can be skipped silently (SkipInvalid), or skip invalid but log a warning (SkipInvalidWithWarning)
TransformProcessRecordReader
to have a transform process applied before being returned.
initialize
Called once at initialization.
param split the split that defines the range of records to read
throws InterruptedException
initialize
Called once at initialization.
param conf a configuration for initialization
param split the split that defines the range of records to read
hasNext
Get the next record
reset
List of label strings
nextRecord
Load the record from the given DataInputStream Unlike {- link #next()} the internal state of the RecordReader is not modified Implementations of this method should not close the DataInputStream
throws IOException if error occurs during reading from the input stream
loadFromMetaData
Load a single record from the given {- link RecordMetaData} instance
Note: that for data that isn’t splittable (i.e., text data that needs to be scanned/split), it is more efficient to load multiple records at once using {- link #loadFromMetaData(List)}
param recordMetaData Metadata for the record that we want to load from
return Single record for the given RecordMetaData instance
throws IOException If I/O error occurs during loading
setListeners
Load multiple records from the given a list of {- link RecordMetaData} instances
param recordMetaDatas Metadata for the records that we want to load from
return Multiple records for the given RecordMetaData instances
throws IOException If I/O error occurs during loading
setListeners
Set the record listeners for this record reader.
close
Closes this stream and releases any system resources associated with it. If the stream is already closed then invoking this method has no effect.
As noted in {- link AutoCloseable#close()}, cases where the close may fail require careful attention. It is strongly advised to relinquish the underlying resources and to internally mark the {- code Closeable} as closed, prior to throwing the {- code IOException}.
throws IOException if an I/O error occurs
setConf
Set the configuration to be used by this object.
getConf
Return the configuration used by this object.
TransformProcessSequenceRecordReader
to be transformed before being returned.
setConf
Set the configuration to be used by this object.
getConf
Return the configuration used by this object.
batchesSupported
Returns a sequence record.
return a sequence of records
nextSequence
Load a sequence record from the given DataInputStream Unlike {- link #next()} the internal state of the RecordReader is not modified Implementations of this method should not close the DataInputStream
throws IOException if error occurs during reading from the input stream
loadSequenceFromMetaData
Load a single sequence record from the given {- link RecordMetaData} instance
Note: that for data that isn’t splittable (i.e., text data that needs to be scanned/split), it is more efficient to load multiple records at once using {- link #loadSequenceFromMetaData(List)}
param recordMetaData Metadata for the sequence record that we want to load from
return Single sequence record for the given RecordMetaData instance
throws IOException If I/O error occurs during loading
initialize
Load multiple sequence records from the given a list of {- link RecordMetaData} instances
param recordMetaDatas Metadata for the records that we want to load from
return Multiple sequence record for the given RecordMetaData instances
throws IOException If I/O error occurs during loading
initialize
Called once at initialization.
param conf a configuration for initialization
param split the split that defines the range of records to read
hasNext
Get the next record
reset
List of label strings
nextRecord
Load the record from the given DataInputStream Unlike {- link #next()} the internal state of the RecordReader is not modified Implementations of this method should not close the DataInputStream
throws IOException if error occurs during reading from the input stream
loadFromMetaData
Load a single record from the given {- link RecordMetaData} instance
Note: that for data that isn’t splittable (i.e., text data that needs to be scanned/split), it is more efficient to load multiple records at once using {- link #loadFromMetaData(List)}
param recordMetaData Metadata for the record that we want to load from
return Single record for the given RecordMetaData instance
throws IOException If I/O error occurs during loading
setListeners
Load multiple records from the given a list of {- link RecordMetaData} instances
param recordMetaDatas Metadata for the records that we want to load from
return Multiple records for the given RecordMetaData instances
throws IOException If I/O error occurs during loading
setListeners
Set the record listeners for this record reader.
close
Closes this stream and releases any system resources associated with it. If the stream is already closed then invoking this method has no effect.
As noted in {- link AutoCloseable#close()}, cases where the close may fail require careful attention. It is strongly advised to relinquish the underlying resources and to internally mark the {- code Closeable} as closed, prior to throwing the {- code IOException}.
throws IOException if an I/O error occurs
NativeAudioRecordReader
Native audio file loader using FFmpeg.
WavFileRecordReader
Wav file loader
ImageRecordReader
Image record reader. Reads a local file system and parses images of a given height and width. All images are rescaled and converted to the given height, width, and number of channels.
Also appends the label if specified (one of k encoding based on the directory structure where each subdir of the root is an indexed label)
TfidfRecordReader
TFIDF record reader (wraps a tfidf vectorizer for delivering labels and conforming to the record reader interface)