fix #526

[nfeybesse]

No description provided.

[Craigacp]

Can you add a test which triggers this bug to the cross entropy tests - https://github.com/tensorflow/java/blob/master/tensorflow-framework/src/test/java/org/tensorflow/framework/losses/CategoricalCrossentropyTest.java. I think this used to work so I worry it's due to a TF upgrade and we didn't catch it with tests.

[nfeybesse]

No, the problem is older, and it is probably the dynamic batch size which triggers the problems. I will try to do a test case

[nfeybesse]

@test
public void testCategoricalCrossEntopyWithDynamicBatchSize() {
try (Graph graph = new Graph()) {
Ops tf = Ops.create(graph);
Operand yPred = tf.placeholder(TFloat32.class, Placeholder.shape(Shape.of(-1, 3)));
Operand yTrue = tf.reshape(tf.constant(new float[] { 1f, 0f, 0f, 0f, 1f, 0f, 0f, 0f, 1f }), tf.array(3, 3));
CategoricalCrossentropy instance = new CategoricalCrossentropy(true);
Operand loss = instance.call(tf, yTrue, yPred);// Throw TFInvalidArgument Exception without fix
try (Session session = new Session(graph); TFloat32 result = (TFloat32) session.runner().feed(yPred, TFloat32.tensorOf(Shape.of(3, 3), DataBuffers.of(new float[] { 1.f, 0.f, 0.f, 0.f, 1.f, 0.f, 0.f, 0.f, 1.f }))).fetch(loss).run().get(0)) {
if (Math.abs(0.5514477f - result.getFloat()) > 0.01)
throw new IllegalStateException("Invalid result :" + result.getFloat());
}
}
}

[nfeybesse]

I am confused because I have the impression that there has not yet been a test carried out by feeding a model with batches of dynamic size. I know from experience that it is largely possible, but that you have to track down a few small bugs. How would you integrate my test so that it would be suitable?

[Craigacp]

Add it next to the other tests for that loss. If there are more issues then let's fix them.

More of the framework was in flight a couple of years ago, but we didn't get all of it merged, so I assume that some of those things were tested in the original codebase before it was broken up into smaller PRs.

[Craigacp]

Thanks