Tensorflow Lite - DataType error: DataType 9 is not recognized in Java

[Whrothus]

I am trying to use a object detection Tensorflow Lite

But When i run

inputTensor = interpreter.GetInputTensor(0);

Error

Java.Lang.IllegalArgumentException: 'DataType error: DataType 9 is not recognized in Java (version 2.0.0)'
interpreter.RunForMultipleInputsOutputs(new Java.Lang.Object[] { input }, outputs);
java.lang.IllegalArgumentException: DataType error: DataType 9 is not recognized in Java (version 2.0.0)
 at Java.Interop.JniEnvironment+InstanceMethods.CallVoidMethod (Java.Interop.JniObjectReference instance, Java.Interop.JniMethodInfo method, Java.Interop.JniArgumentValue* args) [0x0006e] in <8b3b636835d84984ba4604c1f57b1983>:0 
  at Java.Interop.JniPeerMembers+JniInstanceMethods.InvokeAbstractVoidMethod (System.String encodedMember, Java.Interop.IJavaPeerable self, Java.Interop.JniArgumentValue* parameters) [0x00014] in <8b3b636835d84984ba4604c1f57b1983>:0 
  at Xamarin.TensorFlow.Lite.Interpreter.RunForMultipleInputsOutputs (Java.Lang.Object[] inputs, System.Collections.Generic.IDictionary`2[TKey,TValue] outputs) [0x00043] in <4ff04480b90243d7b71dc70e02d1dbfb>:0 

	at org.tensorflow.lite.DataType.fromC(DataType.java:75)
	at org.tensorflow.lite.Tensor.<init>(Tensor.java:363)
	at org.tensorflow.lite.Tensor.fromIndex(Tensor.java:43)
	at org.tensorflow.lite.NativeInterpreterWrapper.getInputTensor(NativeInterpreterWrapper.java:293)
	at org.tensorflow.lite.NativeInterpreterWrapper.run(NativeInterpreterWrapper.java:136)
	at org.tensorflow.lite.Interpreter.runForMultipleInputsOutputs(Interpreter.java:311)

Following code:

var  outputTensor = interpreter.GetOutputTensor(0);

Is no problem

i did try a few other models they work but are toco
I use MLIR

tensorflow/tensorflow#41212 they use Java 2.4

I think its from
https://github.com/tensorflow/tensorflow/blob/98a4bb7a39608275b90bcd221bc51103afadeb10/tensorflow/lite/java/src/main/java/org/tensorflow/lite/Tensor.java#L351

Version update ?

[Whrothus]

Solved
Just Use uint8 :D

converter.inference_input_type = tf.int8
==>
converter.inference_input_type = tf.uint8

[moljac]

My preferred way: Please close the issue yourself if the problem is solved.

Thanks

[Whrothus]

This is NOT fixed i found a solution for me.
Solving it would be Updating the package or Checking when inserting model type if not to use int8 and giving feedback
For now it is a known issue or a Limitation on the package that the main TF supports.

This might seem hard but thx for al the hard work moljac his is a good stuff you do great work

[micosacak]

TF version 1.13.1
Exception in thread "main" java.lang.IllegalArgumentException: DataType 20 is not recognized in Java (version 1.13.1)
at org.tensorflow.DataType.fromC(DataType.java:85)
at org.tensorflow.Tensor.fromHandle(Tensor.java:554)
at org.tensorflow.Session$Runner.runHelper(Session.java:331)
at org.tensorflow.Session$Runner.run(Session.java:264)
at tftest2.main(tftest2.java:56)

I have similar problem.
What is DataType 20? I cannot find explanation for 20?

Here is my problem reported on; https://stackoverflow.com/questions/66654600/datatype-20-is-not-recognized-in-java-version-1-15-0

[micosacak]

System.out.println(inputTensor);

FLOAT tensor with shape [1, 128, 128, 1]

as a result the tensor is (1,128,128,1) shape and float, which is correct!

[Whrothus]

@moljac 28 days
@micosacak Are you doing image recognition ?

[micosacak]

Yes, I am doing image segmentation with Tensorflow.

At the end I want to create a FiJi Plugin. That is why I start to write in Java.

[Whrothus]

Calm down make 1 comment Delete the other please
You can solve this problem to not convert to FLOAT inputtensor [1, 128, 128, 1] (default)
But To Byte input [1,128,128,3]
when you do
converter = tf.lite.TFLiteConverter.from_saved_model(%yourModel%)
Add converter.inference_input_type = tf.uint8
Like example.
converter.txt

Now you got the right model

I am editing https://github.com/jacano/CameraTF
My model SSD [1,320,320,3] runs on my phone att 300ms/frame on real camera input
(I will post my project when its done i am still tweaking)

[micosacak]

@Whrothus I created an example model and maven project and submitted it to this issue: (tensorflow/tensorflow#41212). It can be downloaded from here as well ReportProblem.zip.
I will appreciate if you could let me know, in case you can solve the problem.

[Whrothus]

@micosacak This is a TF lite component so use that
What I did was use a SSD to train a object detection (ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8)
(dataset from https://www.youtube.com/watch?v=Rgpfk6eYxJA&list=WL&index=34&t=1062s )
Converted that to TFlite
And got

got a life camera input feat that +/- 200ms delay on my Huawei P30lite (Small dataset +/- 8h training)
(This is what i got not using Emgu.TF.Lite)
You can change the detection to classification.
This took me +/- 0.5 year in spare time to get it working (start learning Tensorflow to this) (you could do a lot faster)
Now i want to send generate new training records using the new images and telling them if they are not ok (Using FTP) (Hope it works)

This should be updated (my comment 19/2) but you need to use it the right way
Do you have a real dataset ?
Good luck have fun.