Tensorflow in Scala reflection

Question

I am trying to get tensorflow for java to work on Scala. I am use the tensorflow java library without any wrapper for Scala.

At sbt I have:

If I run the HelloWord found here, it WORKS fine, with the Scala adaptations:

import org.tensorflow.Graph
import org.tensorflow.Session
import org.tensorflow.Tensor
import org.tensorflow.TensorFlow


val g = new Graph()
val value = "Hello from " + TensorFlow.version()
val t = Tensor.create(value.getBytes("UTF-8"))
// The Java API doesn't yet include convenience functions for adding operations.
g.opBuilder("Const", "MyConst").setAttr("dtype", t.dataType()).setAttr("value", t).build();

val s = new Session(g)
val output = s.runner().fetch("MyConst").run().get(0)

However, if I try to use Scala reflection to compile the function from a string, it DOES NOT WORK. Here is the snippet I used to run:

import scala.reflect.runtime.{universe => ru}
import scala.tools.reflect.ToolBox
val fnStr = """
    {() =>
      import org.tensorflow.Graph
      import org.tensorflow.Session
      import org.tensorflow.Tensor
      import org.tensorflow.TensorFlow

      val g = new Graph()
      val value = "Hello from " + TensorFlow.version()
      val t = Tensor.create(value.getBytes("UTF-8"))
      g.opBuilder("Const", "MyConst").setAttr("dtype", t.dataType()).setAttr("value", t).build();

      val s = new Session(g)

      s.runner().fetch("MyConst").run().get(0)
    }
    """
val mirror = ru.runtimeMirror(getClass.getClassLoader)
val tb = mirror.mkToolBox()
var t = tb.parse(fnStr)
val fn = tb.eval(t).asInstanceOf[() => Any]
// and finally, executing the function
fn()

Here simplified build.sbt to reproduce the error above:

lazy val commonSettings = Seq(
    scalaVersion := "2.12.10",

    libraryDependencies ++= {
      Seq(
                  // To support runtime compilation
        "org.scala-lang" % "scala-reflect" % scalaVersion.value,
        "org.scala-lang" % "scala-compiler" % scalaVersion.value,

        // for tensorflow4java
        "org.tensorflow" % "tensorflow" % "1.15.0",
        "org.tensorflow" % "proto" % "1.15.0",
        "org.tensorflow" % "libtensorflow_jni" % "1.15.0"

      )
    }
)

lazy val `test-proj` = project
  .in(file("."))
  .settings(commonSettings)

When running the above, for example with sbt console, I get the following error and stack trace:

java.lang.NoSuchMethodError: org.tensorflow.Session.runner()Lorg/tensorflow/Session$$Runner;
  at __wrapper$1$f093d26a3c504d4381a37ef78b6c3d54.__wrapper$1$f093d26a3c504d4381a37ef78b6c3d54$.$anonfun$wrapper$1(<no source file>:15)

Please ignore the memory-leaks that the previous code has given that no resources context (to close()) is used

Why do you need to compile the code from a string with ToolBox? — Dmytro Mitin, Mar 25 '20 at 23:32
Think of a serveless or Function As a Service (FaaS) scenario, for example. — aitorhh, Mar 26 '20 at 10:46
Maybe as a workaround you can programmatically write this code string to a text file, compile the file and run compiled class file. Does this work for you? — Dmytro Mitin, Mar 29 '20 at 13:46
Thanks @DmytroMitin, it is a valid workaround. I am also considering tensorflow for scala, but I might found similar issue — aitorhh, Mar 30 '20 at 10:12

Dmytro Mitin · Accepted Answer · 2020-11-26T13:04:00.700

The thing is in this bug appearing in combination of reflective compilation and Scala-Java interop

https://github.com/scala/bug/issues/8956

Toolbox can't typecheck a value (s.runner()) of path-dependent type (s.Runner) if this type comes from Java non-static inner class. And Runner is exactly such class inside org.tensorflow.Session.

You can run the compiler manually (similarly to how Toolbox runs it)

import org.tensorflow.Tensor
import scala.reflect.internal.util.{AbstractFileClassLoader, BatchSourceFile}
import scala.reflect.io.{AbstractFile, VirtualDirectory}
import scala.reflect.runtime
import scala.reflect.runtime.universe
import scala.reflect.runtime.universe._
import scala.tools.nsc.{Global, Settings}

val code: String =
  """
    |import org.tensorflow.Graph
    |import org.tensorflow.Session
    |import org.tensorflow.Tensor
    |import org.tensorflow.TensorFlow
    |
    |object Main {
    |  def foo() = () => {
    |      val g = new Graph()
    |      val value = "Hello from " + TensorFlow.version()
    |      val t = Tensor.create(value.getBytes("UTF-8"))
    |      g.opBuilder("Const", "MyConst").setAttr("dtype", t.dataType()).setAttr("value", t).build();
    |
    |      val s = new Session(g)
    |
    |      s.runner().fetch("MyConst").run().get(0)
    |  }
    |}
""".stripMargin

val directory = new VirtualDirectory("(memory)", None)
val runtimeMirror = createRuntimeMirror(directory, runtime.currentMirror)
compileCode(code, List(), directory)
val tensor = runObjectMethod("Main", runtimeMirror, "foo").asInstanceOf[() => Tensor[_]]
tensor() // STRING tensor with shape []

def compileCode(code: String, classpathDirectories: List[AbstractFile], outputDirectory: AbstractFile): Unit = {
  val settings = new Settings
  classpathDirectories.foreach(dir => settings.classpath.prepend(dir.toString))
  settings.outputDirs.setSingleOutput(outputDirectory)
  settings.usejavacp.value = true
  val global = new Global(settings)
  (new global.Run).compileSources(List(new BatchSourceFile("(inline)", code)))
}

def runObjectMethod(objectName: String, runtimeMirror: Mirror, methodName: String, arguments: Any*): Any = {
  val objectSymbol = runtimeMirror.staticModule(objectName)
  val objectModuleMirror = runtimeMirror.reflectModule(objectSymbol)
  val objectInstance = objectModuleMirror.instance
  val objectType = objectSymbol.typeSignature
  val methodSymbol = objectType.decl(TermName(methodName)).asMethod
  val objectInstanceMirror = runtimeMirror.reflect(objectInstance)
  val methodMirror = objectInstanceMirror.reflectMethod(methodSymbol)
  methodMirror(arguments: _*)
}

def createRuntimeMirror(directory: AbstractFile, parentMirror: Mirror): Mirror = {
  val classLoader = new AbstractFileClassLoader(directory, parentMirror.classLoader)
  universe.runtimeMirror(classLoader)
}

dynamically parse json in flink map

Dynamic compilation of multiple Scala classes at runtime

How to eval code that uses InterfaceStability annotation (that fails with "illegal cyclic reference involving class InterfaceStability")?

Thanks for the answer! While your answer does work, there it is needed to refer to the Main.foo() function, and in the question we have directly the lambda function. For sure nothing critical! This is definetely the way to go — aitorhh, Dec 01 '20 at 13:44
@aitorhh Well, in Scala 2 a lambda can't be top-level, it should be inside some method or constructor of a class or object. Just Toolbox hides this wrapping from your eyes. While running compiler manually I did that a little more explicitly. If you prefer to hide that you can do like in https://stackoverflow.com/questions/53976254/how-to-eval-code-that-uses-interfacestability-annotation-that-fails-with-illeg — Dmytro Mitin, Dec 01 '20 at 13:52

score 1 · Answer 2 · answered Dec 01 '20 at 14:45

As Dmytro pointed out on his answer, it is not possible using toolbox. And he pointed out to another answer (How to eval code that uses InterfaceStability annotation (that fails with "illegal cyclic reference involving class InterfaceStability")?). I think there is a neat solution by just replace the Compiler class defined in the previous, and replacing the Toolbox for that Compiler class.

In that case, the final snippet will look like:

import your.package.Compiler
val fnStr = """
    {() =>
      import org.tensorflow.Graph
      import org.tensorflow.Session
      import org.tensorflow.Tensor
      import org.tensorflow.TensorFlow

      val g = new Graph()
      val value = "Hello from " + TensorFlow.version()
      val t = Tensor.create(value.getBytes("UTF-8"))
      g.opBuilder("Const", "MyConst").setAttr("dtype", t.dataType()).setAttr("value", t).build();

      val s = new Session(g)

      s.runner().fetch("MyConst").run().get(0)
    }
    """
val tb = new Compiler() // this replaces the mirror and toolbox instantiation
var t = tb.parse(fnStr)
val fn = tb.eval(t).asInstanceOf[() => Any]
// and finally, executing the function
println(fn())

And just for completion, copy/paste from the solution at this answer:

  class Compiler() {
    import scala.reflect.internal.util.{AbstractFileClassLoader, BatchSourceFile}
    import scala.reflect.io.{AbstractFile, VirtualDirectory}
    import scala.reflect.runtime
    import scala.reflect.runtime.universe
    import scala.reflect.runtime.universe._
    import scala.tools.nsc.{Global, Settings}
    import scala.collection.mutable
    import java.security.MessageDigest
    import java.math.BigInteger
       
    val target  = new VirtualDirectory("(memory)", None)
       
    val classCache = mutable.Map[String, Class[_]]()
       
    private val settings = new Settings()
    settings.deprecation.value = true // enable detailed deprecation warnings
    settings.unchecked.value = true // enable detailed unchecked warnings
    settings.outputDirs.setSingleOutput(target)
    settings.usejavacp.value = true
       
    private val global = new Global(settings)
    private lazy val run = new global.Run
       
    val classLoader = new AbstractFileClassLoader(target, this.getClass.getClassLoader)
       
    /**Compiles the code as a class into the class loader of this compiler.
      * 
      * @param code
      * @return
      */
    def compile(code: String) = {
      val className = classNameForCode(code)
      findClass(className).getOrElse {
        val sourceFiles = List(new BatchSourceFile("(inline)", wrapCodeInClass(className, code)))
        run.compileSources(sourceFiles)
        findClass(className).get
      } 
    }   
       
    /** Compiles the source string into the class loader and
      * evaluates it.
      * 
      * @param code
      * @tparam T
      * @return
      */
    def eval[T](code: String): T = {
      val cls = compile(code)
      cls.getConstructor().newInstance().asInstanceOf[() => Any].apply().asInstanceOf[T]
    }  
        
    def findClass(className: String): Option[Class[_]] = {
      synchronized {
        classCache.get(className).orElse {
          try {
            val cls = classLoader.loadClass(className)
            classCache(className) = cls
            Some(cls)
          } catch {
            case e: ClassNotFoundException => None
          }
        }
      } 
    }   
  
    protected def classNameForCode(code: String): String = {
      val digest = MessageDigest.getInstance("SHA-1").digest(code.getBytes)
      "sha"+new BigInteger(1, digest).toString(16)
    }   
  
    /*  
     * Wrap source code in a new class with an apply method.
     */ 
   private def wrapCodeInClass(className: String, code: String) = {
     "class " + className + " extends (() => Any) {\n" +
     "  def apply() = {\n" +
     code + "\n" +
     "  }\n" +
     "}\n"
   }    
  }

Tensorflow in Scala reflection

2 Answers2

Linked