DataVec has been migrated to a sub-module of the mono-repository here: https://github.com/deeplearning4j/deeplearning4j All future development will continue at that repository, which should be used for all issues and pull requests.
DataVec is an Apache 2.0-licensed library for machine-learning ETL (Extract, Transform, Load) operations. DataVec's purpose is to transform raw data into usable vector formats that can be fed to machine learning algorithms. By contributing code to this repository, you agree to make your contribution available under an Apache 2.0 license.
Data handling is sometimes messy, and we believe it should be distinct from high-performance algebra libraries (such as nd4j or Deeplearning4j).
DataVec allows a practitioner to take raw data and produce open standard compliant vectorized data (svmLight, etc) quickly. Current input data types supported out of the box:
- CSV Data
- Raw Text Data (Tweets, Text Documents, etc)
- Image Data
- LibSVM
- SVMLight
- MatLab (MAT) format
- JSON, XML, YAML, XML
Datavec draws inspiration from a lot of the Hadoop ecosystem tools, and in particular accesses data on disk through the Hadoop API (like Spark does), which means it's compatible with many records
DataVec also includes sophisticated functionality for feature engineering, data cleaning and data normalization both for static data and for sequences (time series). Such operations can be executed on Apache Spark using DataVec-Spark.
Apart from obviously providing readers for classic data formats, DataVec also provides an interface. So if you wanted to ingest specific custom data, you don't have to do the whole pipeline, you just have to do the very first step. You describe through the API how your data fits into a common format that complies with the interface, in this case, DataVec will return a list of Writables for each record. You'll find more detail on the API in the corresponding module.
Another thing you can do with DataVec is data cleaning functionality. Instead of having clean ready-to-go data, say you start with maybe data in different forms or from different sources. You might need to do sampling, filtering, or several of all those incredibly messy ETL tasks that you need to prepare data in the real world. DataVec offers filters and transformations that help with curating, preparing and massaging your data. It leverages Apache Spark to do this at scale.
Finally, DataVec tracks a schema for your columnar data, across all transformations. This schema is actively checked
through probing, and DataVec will raise exceptions if your data does not match the schema. You can specify filters as
well: you can attach a regular expression to an input column of type String, for example, and DataVec will only keep
data that matches this filter
Distributed treatment through Apache Spark is entirely optional, including running Spark in local-mode (where your cluster is emulated with multi-threading) when necessary. Datavec aims to abstract away from the actual execution and create at compile time, a logical set of operations to execute. While we have some code that uses Spark, we do not want to be locked into a single tool, and using Apache Flink or Beam are possibilities, on which we would welcome collaboration.
Examples for using DataVec are available here: https://github.com/deeplearning4j/dl4j-examples
We have a lot on the pipeline, and even more we'd love to receive contributions. We want to support representing data as more than a collection of simple types ("writables"), and rather as binary data — that will help with GC pressure across our pipelines and fit better with media-based uses cases, where columnar data is not essential. We also expect it will streamline a lot of the specialized operations we now do on primitive types.
With that being said, an area that could welcome a first contribution is the implementations of the RecordReader
interface, since this is relatively self-contained. Of note, to support most of the distributed file formats of the
Hadoop ecosystem, we use Apache Camel. Camel supports
a pluggable DataFormat to allow messages to be marshalled to and from
binary or text formats to support a kind of Message Translator.
Another area that is relatively self-contained is transformations, where you might find a filter or data munging operation that has not been implemented yet, and provide it in a self-contained way.
It's often useful to have an idea of which maintainers to contact to get information on a particular part of the code, including reviewing your pull requests, or asking questions on our gitter channel. For this you can use the following, indicative mapping:
RecordReaderimplementations: @saudet and @agibsonccc- Transformations and their API: @agibsonccc and @AlexDBlack
- Spark and distributed processing: @AlexDBlack, @agibsonccc and @huitseeker
- Native formats, geodata: @saudet
-
Check for open issues, or open a new issue to start a discussion around a feature idea or a bug.
-
If you feel uncomfortable or uncertain about an issue or your changes, feel free to contact us on Gitter using the link above.
-
Fork the repository on GitHub to start making your changes.
-
Write a test, which shows that the bug was fixed or that the feature works as expected.
-
Note the repository follows the Google Java style with two modifications: 120-char column wrap and 4-spaces indentation. You can format your code to this format by typing
mvn formatter:formatin the subproject you work on, by using thecontrib/formatter.xmlat the root of the repository to configure the Eclipse formatter, or by using the INtellij plugin. -
Send a pull request, and bug us on Gitter until it gets merged and published.
- Downloading the latest jar from https://projectlombok.org/download
- Double click the jar to install the plugin for Eclipse
- Clone datavec to your system
- Import the project as a maven project
- You will also need clone and build ND4J and libnd4j