This post is the first in a series of community show case posts. Here we aggregate and highlight notable posts by members of the broader Deeplearning4J community. Where possible we aim to reproduce those posts here.
What would be a better topic for the first post of this series than a setup guide? This guide shows you how to set up your system in order to be able to use DL4J and run a few examples.
- Why Java?
- Deep Learning for the JVM — Eclipse Deeplearning4J (DL4J)
- Set up environment for the first time with DL4J
- Run DL4J Examples in 3 steps
Below are three of the fundamental reasons to use Java programming language for Machine Learning operations.
Write Once, Run Anywhere
Java programming language is platform-agnostic which can runs on Linux, Mac and Windows. This feature of “Write once, run anywhere” is made possible by Java Virtual Machine (JVM). With Java compiler generates Java binary code (bytecode) from Java source code, the generated bytecode can run on JVM regardless of the machines.
High Performance and Automatic Memory Management
Java is built to deliver high performance with Just-In-Time (JIT) compiler. JIT compiler transforms Java bytecode into native instructions at runtime, performing optimization in the process. Other than that, Garbage Collection makes Java efficient by performing automatic memory management.
The Standard of the Industry
The use of Java whether on chips, devices, or software packages has become the industry standard practice in production. Developing Deep Learning algorithms using Java takes into the consideration of actual production environment, not just research in the lab. Lots of big companies rewrote machine learning code to a language suited to the system during the stage of AI models deployment, resulted in delay entry to the market.
Deep Learning for the JVM: Eclipse DeepLearning4J
Eclipse DeepLearning4J (DL4J) is an open source, JVM-based Deep Learning framework. DL4J provides a suite of tools for building production-grade Deep Learning applications. DL4J also supports the integration with Apache Spark and Hadoop, allowing training and inference on CPU or GPU cluster to further accelerate machine learning workloads.
DL4J comprised of a suite of tools such as DataVec, ND4J, LibND4J, RL4J, and others. The modules come together to support Deep Learning operations. You can read more about the complete list of the sub-modules here.
Figure 1 shows how DeepLearning4J works from neural network modelling to “close to the metal”, controlling the whole software stack. First of all, DataVec serves as a vectorization tool for data of multiple sources and formats. DeepLearning4J sub-module comes with functionalities to build from multi-layer networks to computation graphs. It also allows the import of Keras and Tensorflow models.
The backends of DL4J is ND4J (think of it as Numpy for the JVM), it is a linear algebra library with switchable backends of either using CPUs or GPUs. The ability to leverage LibND4J (written in C++) for hardware acceleration largely contributed by the existence of JavaCPP, acting as a bridge between ND4J and LibND4J.
Note: JavaCPP is not a module out of Eclipse DeepLearning4J. It is an independent software distribution maintained by Bytedeco.
Set up environment for the first time with DL4J
Let’s walk through the workflow as shown in Figure 2 to install and configure the paths for these prerequisites:
- Java — Programming Language
- Apache Maven — Dependency Management Tool
- Git — Version Control System
- Intellij IDEA — Integrated Development Environment
If you already have Steps 1–4 fulfilled, jump to next section Run DL4J Examples in 3 Steps.
For Windows 10
Select and download the file with the product name “Windows x64” from here. Make sure you are choosing 64-Bit version of Java. You might need to create Oracle account to proceed with the download.
JAVA_HOME Variable Configuration
To set JAVA_HOME path:
- Search for Edit the system environment variables
- Click the Environment Variables button.
- Under System Variables, click New.
- In the Variable Name field, enter JAVA_HOME
- In the Variable Value name, enter your JDK installation path.
For Ubuntu 18.04 LTS
Open a terminal and install with command
sudo apt install openjdk-8-jdk
Select and download the file(*.dmg) for Mac OS from here.
For three of the systems above, verify that Java is working right after the installation with the command as below:
Installation of Java Deployment Toolkit (JDK) comes with the associated Java Runtime Environment (JRE) and JavaFX SDK. These will be installed and integrated into the JDK directory structure.
In DL4J projects, Apache Maven is mandatory for process such as clean, build, package and install while managing the package dependencies and versions. As your Java projects get more complex, you will be glad that you use Maven instead of using the native “javac”, “java -jar” command line approaches.
Follow the instruction from this link to install Maven on your system.
Verify that Maven is successfully installed using the command below:
If you are new to Maven, check out this link: Maven in 5 minutes.
Git is the mostly widely used and the standard for version controlling system. Follow the instructions on this link to install the latest Git on your system.
Verify that Git is successfully installed with the command:
Intellij IDEA is the Integrated Development Environment (IDE) preferably used for DL4J projects. Alternatively, you can also use other IDEs such as Eclipse and Netbeans, but expect the process to be much more complex and error-prone (especially with dealing with dependencies).
For Windows and Mac
Download and install Intellij Community Version from here.
The recommended way would be installing Intellij through Software Center.
Run DL4J Examples in 3 Steps
After you have the prerequisites all set up, lets proceed with cloning the dl4j-examples repository to explore more. This repository contains a comprehensive examples on configuring neural networks on DL4J for various use cases.
Enter the following to your terminal/command line tool.
git clone https://github.com/deeplearning4j/dl4j-examples.git
Open the repository in Intellij by clicking on the repository.
After importing the project, resolving Maven dependencies might take awhile (10 minutes to more!) depending your networking speed. Come back after a cup of coffee. You will need to wait till the progress bar (normally at the bottom right corner) to disappear before proceeding to the next step.
You can choose to run any examples in the directories. For illustration, I ran ImageDrawer.java which learns how to draw an image mimicking an input image. You can find the file path to this example through Figure 14.
The result is pretty promising as you can see. The silhouette of Mona Lisa painting is forming out as shown in the generated image (left) of Figure 15.