Tensorflow java fails when a variable is hidden from gradients using tf.stopGradient.

Variables without gradients are used to have persistent state between runs. In my case, I want to have some customizable values in the graph. With a constant, these values cannot be changed. For instance:

For instance, I want to normalize input values so that their square deviation is (more or less) 1. So one introduces a reasonable value of square deviation via a Tensorflow variable defined with

var signalVar = tf.withName("signalVar").variable(tf.constant(1.f));

it is initialized as 1, but can be modified with variable assignment (omitted here). Then use it to divide the input signal. But we don't want to train it.

var signalNormalized = tf.math.div(signal, tf.stopGradient(signalNormalized));

Then, when calling optimizer.minimize an IllegalStateException "close() has been called on the Graph this Operation was a part of".

var optimizer = new Adam(graph, learningRate, betaOne, betaTwo, epsilon);
optimizer.minimize(loss, "train"); // IllegalStateException throw here.

The reason is that when the optimizer code scans for variables, it checks the gradients and their data types, but a variable protected with stopGradient() has no gradient TF_Operation initialized.