Problems with Variable Scope

[mandar2812]

After migrating DynaML code from tf-0.2.4 to tf-0.4.1 or tf-0.4.2-SNAPSHOT

Accessing variables becomes buggy.

When using the tf.variable() method

tf.variable[D]("Linear_1/Weights", shape = Shape(-1, 10), reuse = ReuseExistingOnly)

I see the following exception stack trace

java.lang.IllegalArgumentException: Variable 'Loss/ProbWeightedTS/Loss/ProbWeightedTS//Linear_1/Weights' does not exist, but variable scope re-use was set to 'ReuseExistingOnly'.
  org.platanios.tensorflow.api.ops.variables.VariableStore.getVariable(VariableStore.scala:99)
  org.platanios.tensorflow.api.ops.variables.VariableScope.$anonfun$getVariable$1(VariableScope.scala:105)
  scala.util.DynamicVariable.withValue(DynamicVariable.scala:58)
  org.platanios.tensorflow.api.ops.Op$.nameScope(Op.scala:2067)
  org.platanios.tensorflow.api.ops.variables.VariableScope.getVariable(VariableScope.scala:105)
  org.platanios.tensorflow.api.ops.variables.Variable$.getVariable(Variable.scala:508)
  org.platanios.tensorflow.api.ops.variables.package$API.variable(package.scala:86)
  org.platanios.tensorflow.api.ops.variables.package$API.variable$(package.scala:80)
  org.platanios.tensorflow.api.package$tf$.variable(package.scala:294)
  io.github.mandar2812.dynaml.tensorflow.layers.L2Regularization.$anonfun$forwardWithoutContext$1(L2Regularization.scala:45)

The string Loss/ProbWeightedTS is the name of the Loss function, in earlier versions of tf-scala (until 0.2.4) this was appended to the variable name only once, now it seems the VariableScope is being appended twice Loss/ProbWeightedTS/Loss/ProbWeightedTS//Linear_1/Weights when using tf.variable()

[eaplatanios]

@mandar2812 Was this resolved or is it still an issue that would need to be fixed in TF Scala?

[mandar2812]

@eaplatanios No this does not seem to be an issue in TF Scala :)
After looking in the API, I was able to use tf.variableScope() and tf.nameScope() to my advantage.
What I currently have as a stop gap for #88 is using these two methods to access variables by name (with tf.ReuseExistingOnly).