tf.variable with initial value does not work in eager mode

[rnett]

As the title says, tf.variable with an initial value does not work in eager mode. I'm not sure if it is supposed to, as tf.Variable does in the python api but apparently didn't at first. Interestingly, it's not the variable op that fails, it's the assign:

cannot compute Assign as input #0(zero-based) was expected to be a int32_ref tensor but is a int32 tensor
org.tensorflow.exceptions.TFInvalidArgumentException: cannot compute Assign as input #0(zero-based) was expected to be a int32_ref tensor but is a int32 tensor
	at org.tensorflow.internal.c_api.AbstractTF_Status.throwExceptionIfNotOK(AbstractTF_Status.java:87)
	at org.tensorflow.EagerOperationBuilder.execute(EagerOperationBuilder.java:273)
	at org.tensorflow.EagerOperationBuilder.build(EagerOperationBuilder.java:66)
	at org.tensorflow.EagerOperationBuilder.build(EagerOperationBuilder.java:55)
	at org.tensorflow.op.core.Assign.create(Assign.java:97)
	at org.tensorflow.op.core.Helpers.createVariableWithInit(Helpers.java:51)
	at org.tensorflow.op.Ops.variable(Ops.java:7520)

[Craigacp]

Yeah, that's documented in the javadoc:

java/tensorflow-core/tensorflow-core-api/src/main/java/org/tensorflow/op/core/Helpers.java

Line 40 in 8c67a5a

* Only supported on Graph sessions as the {@link org.tensorflow.op.core.Assign} op

I couldn't figure out an alternative at the time and needed it for the optimizers work. We should at minimum fix it so it throws a sensible looking exception.

[rnett]

Could be done easily enough as part of #154, that's what I was originally going to suggest before I realized it was supported in python eager mode.

I assume the codegen doesn't support custom javadocs on the ops methods (for putting the warning on tf.variable)?