Ok, getting the hang of it. Is this the idiomatic way?
val g = new Graph()
val ops = Ops.create(g)
import ops.*
val a = constant(3.0)
val b = constant(2.0)
val x = placeholder(classOf[TFloat64])
val y = placeholder(classOf[TFloat64])
val ax = math.mul(a, x)
val by = math.mul(b, y)
val z = math.add(ax, by)
val session = new Session(g)
val r = session.runner()
val tensor = r.fetch(z)
.feed(x, TFloat64.scalarOf(3.0))
.feed(y, TFloat64.scalarOf(6.0))
.run
.get(0)
tensor match
case tf: TFloat64 => println(tf.getDouble()) // 21.0
case other => println(s"OTHER: $other")
That tutorial is for the old 1
1.x API. The tests in the TensorFlow Java repo are the most up to date way of doing things. Note at the moment there's an issue with gradient determinism when building models for training that causes incorrect outputs occasionally.
We are working on a higher level API, but it's not all available yet
Is it possible to use keras API, e.g. https://www.tensorflow.org/tutorials/generative/dcgan ? or do I have to translate that back to plain TF ?
Ok, I'm trying to update an older keras wrapper project, it uses the old Java API. What this the new way of writing
Tensors.create(float[][])?
Tensor.of(TFloat32.class, Shape.of(javaArray.length, javaArray[0].length), data -> StdArrays.copyTo(javaArray, data)) I guess ... ? Is there a utility function that does this?
I'm stuck with these:
public class Ones {
public <T extends TNumber> Operand<T> initialize(Ops tf, Operand<TInt32> shape, Class<T> dtype) {
return tf.fill(shape, tf.constant(1.0f, dtype)); // no such method
}
}How can I convert from a float to a TNumber? In Scala this would be a type class instance TNumbers[T <: TNumber] { def fromFloat(f: Float): Operand[T] }, but I don't know how to express that in Java. Should I change Class<T> for something that has factory capabilities?
If you can avoid completely the use of standard Java arrays and only rely on the
NdArray hierarchy from the start, you’ll have better performances and memory usage
Hi, I have a question concerning saving/exporting/loading a model for deployment. I think I want something like a frozen graph of my model i.e. one single pb file. I used tf.io.write_graph in python to write my model into a pb file. However when trying to load it in Java via SavedModelLoader.loader(pb_file_directory) I get an exception: "Could not find meta graph def matching supplied tags: { serve }. To inspect available tag-sets in the SavedModel, please use the SavedModel CLI:
saved_model_cli". And saved_model_cli tells me that "SavedModel file does not exist" in that directory. I'm confused and my question is: how is this usually done and what works for TF java?
If you've got a frozen graph you can load that directly by reading it into a GraphDef and supplying that to the Graph before making a session. That's not how TF 2 (or even later versions of TF 1) do deployment, they use SavedModels which is a folder containing a graph pb and a checkpoint pb of the variables.
TF-Java uses SavedModelBundle to load in saved models.
Trying to translate this tutorial. Python:
noise = tf.random.normal([1, 100])How do I write that in Java/Scala? I try
val noise = tf.random.randomStandardNormal(Shape.of(1, 100), classOf[TFloat32])But this is not an Operand[? <: TNumber]. How I get from Shape to that?
Is
tf.constant(Array(1, 100)) correct?
You can look at how we use them here - https://github.com/tensorflow/java/blob/master/tensorflow-framework/src/main/java/org/tensorflow/framework/initializers/RandomNormal.java
Hi there. I've a question again: I want to use mask_rcnn_inception_resnet_v2_1024x1024_coco17_gpu-8 from the TF2 model zoo for semantic image segmantation. I trained the CNN with python tools and now I'm trying to get the masks in our JVM inference implementation. I wonder if I can use the implementation for the "normal" FasterRCNN resulting to BBoxes. Are the masks available in the outputTensorMap? There is a key "mask_prediction" with a Tensor of shape [100, 4, 33, 33] which looks promising. Am I on the right track? How is it thought to access the masks in the java world?
If you build a graph and want to train with automatic BP, the optimizers I’ve pointed at will do the job for you, check here for an example.
Gradient tapes are used to calculate loss in eager mode (i.e. without graphs). If you use a graph, you are good to go.
Is it ok to ask questions regarding the TF java api here?
I'm not sure if you remember, but you posted some code that basically did some pre-processing on the buffered image bytes before putting them into the Tensor.
could be in this same thread
so other users looking at this thread and having the same questions as you will see the answer (hopefully :) )