2017-11-06 Update: Deeplearning4J

Not much on the Clojure API front.

I did look at Deeplearning4J, a deep learning framework written in Java. I cloned their repo of examples, and I got some of them working with gradle files, and learned a bit more about Gradle in the process. You can use a GPU, or a CPU. No messing with CUDA if you don’t want to.

There is even a Clojure wrapper for Deeplearning4J as well.

I had a post about using CPUs and GPUs from JVM languages for math/science/AI a while back. Since then, I have gotten both Neanderthal and Deeplearning4J to work on my system. I did mention Bytedeco, which is used by Deeplearning4J and a few other libraries. I looked at some of the Javadoc, and they use a LOT of static methods (see here for an example). I may leave Bytedeco for others and use some of the higher-level libraries, like Deeplearning4J. I think a lot of static methods is not idiomatic Java. I might consult “Effective Java” later for a second opinion.

I am going to include below one of the build.gradle files that I used to get some of the Deeplearning4J examples to work.

You’re welcome.

apply plugin: 'java'
apply plugin: 'maven'
// apply plugin: 'groovy'
apply plugin: 'codenarc'
// apply plugin: 'application'

repositories {
  mavenCentral()
  jcenter()
}

// this has to be at the top so runJava can read mainClassProp
ext {
    dl4jVersion        = '0.9.1'
    nd4jVersion        = '0.9.1'
    jacksonVersion     = '2.6.6'
    logbackVersion     = '1.1.7'
    nd4jBackend        = 'nd4j-native-platform'
    scalaBinaryVersion = '2.11'
    dl4jSparkVersion   = '0.9.1_spark_1'
    awsSdkVersion      = '1.11.109'
    jcommonVersion     = '1.0.23'
    mainClassProp      = 'org.deeplearning4j.transferlearning.vgg16.dataHelpers.FeaturizedPreSave'
}

/*
'org.deeplearning4j.emr.EmrSparkExample'
'org.deeplearning4j.transferlearning.vgg16.FitFromFeaturized'
'org.deeplearning4j.transferlearning/vgg16.dataHelpers.FeaturizedPreSave'
'org.deeplearning4j.rnn.SparkLSTMCharacterExample'
'org.deeplearning4j.stats.TrainingStatsExample'
'org.deeplearning4j.mlp.MnistMLPExample'
'org.deeplearning4j.mlp.MnistMLPDistributedExample'
*/

allprojects {
    dependencies {
        // testCompile 'junit:junit:4.10'
        // testCompile 'org.mockito:mockito-all:[1.8.5,)'
        // From Maven, it's GroupId:ArtifactId:Version 
    
        // logging
        // Using sfl4j allows us to replace different loggers later, if desired
        compile group: 'org.slf4j', name:'slf4j-api', version: '1.7.2'
        // sfl4j is just a facade for loggers. To get logging to work, tie it in with an implementation, e.g., logback
        compile group: 'ch.qos.logback', name:'logback-classic', version: "${logbackVersion}"
        compile group: 'ch.qos.logback', name:'logback-core',    version: "${logbackVersion}"
        
        compile "org.deeplearning4j:deeplearning4j-zoo:${dl4jVersion}"
        compile "com.fasterxml.jackson.core:jackson-databind:${jacksonVersion}"
        compile "com.fasterxml.jackson.core:jackson-annotations:${jacksonVersion}"
        compile "com.fasterxml.jackson.module:jackson-module-scala_2.11:${jacksonVersion}"

        compile "org.nd4j:${nd4jBackend}:${nd4jVersion}"

        compile "org.deeplearning4j:dl4j-spark_${scalaBinaryVersion}:${dl4jSparkVersion}"

        compile "org.deeplearning4j:dl4j-spark-parameterserver_${scalaBinaryVersion}:${dl4jSparkVersion}"

        compile "com.beust:jcommander:${jcommonVersion}"

		compile "com.amazonaws:aws-java-sdk-emr:${awsSdkVersion}"
		// scope: provided
		
		compile "com.amazonaws:aws-java-sdk-s3:${awsSdkVersion}"
		// scope provided
        // compile "org.nd4j:${nd4jBackend}:${nd4jVersion}"
        // for CUDA
        // compile "org.nd4j:nd4j-cuda-7.5-platform:${nd4jVersion}"
    }
}

task runJava( type: JavaExec ) {
    println( "-- Groovy version: " + groovy.lang.GroovySystem.getVersion() )
    dependsOn classes
    description = 'Run gradle.sample.SampleApp'
    // Java main class to execute.
    if ( rootProject.hasProperty( "mainClass" ) ) {
        main = mainClass
    } else {
        main = mainClassProp
    }
    // We need to set the classpath.
    classpath sourceSets.main.runtimeClasspath
    // Extra options can be set.
    systemProperty 'sysProp', 'notUsed'
    println System.properties['iterations'] 
    // systemProperty 'iterations', System.getProperty( 'iterations' )
    systemProperty 'iterations', System.properties[ 'iterations' ] 
    // you can use this if nothing is set via command line
    // systemProperties = System.getProperties()
    jvmArgs '-server'
    // We can pass arguments to the main() method
    // of gradle.sample.SampleApp.
    if ( rootProject.hasProperty( "mainArgs" ) ) {     
        args = mainArgs.tokenize() // args wants a List, so split by spaces if need be
    } 
}

Image from the Trebizond Gospels, a 12th-century Byzantine manuscript housed at The Walters Art Museum, manuscript information here, image from World Document Library, image assumed allowed under Fair Use.