Cannot use a simple RF model trained on the Iris dataset

[codepictor]

Hello!

First of all, thank you for developing this really promising tool!

I have studied the examples in the separate repo: https://github.com/JetBrains-Research/kinference-examples. Now I would like to run RandomForestClassifier trained on the Iris dataset, as demonstrated in the following ONNX tutorial: https://onnx.ai/sklearn-onnx/index.html. This code successfully produces the file called "rf_iris.onnx".

Then, I try to run "rf_iris.onnx" in Kotlin with the help of KInference (KIEngine is used to load the model):

val input = FloatNDArray(shape=intArrayOf(1, 4)) { it.toFloat() }

val output = runBlocking { model.predict(listOf(input.asTensor("X"))) }

However, I get ArrayIndexOutOfBoundsException:

Index 2280 out of bounds for length 2277
java.lang.ArrayIndexOutOfBoundsException: Index 2280 out of bounds for length 2277
at io.kinference.trees.Sum.accept(Aggregator.kt:27)
at io.kinference.trees.SingleModeTreeEnsemble.applyEntry(SingleModeTreeEnsemble.kt:30)
at io.kinference.trees.SingleModeTreeEnsemble.applyEntry$default(SingleModeTreeEnsemble.kt:19)
at io.kinference.core.operators.ml.trees.KICoreTreeEnsemble.execute(KICoreTreeEnsemble.kt:40)
at io.kinference.core.operators.ml.TreeEnsembleClassifierVer1.apply(TreeEnsembleClassifier.kt:132)
at io.kinference.operator.Operator.applyWithCheck(Operator.kt:174)
at io.kinference.graph.Graph$execute$3$1.invokeSuspend(Graph.kt:224)
at io.kinference.graph.Graph$execute$3$1.invoke(Graph.kt)
at io.kinference.graph.Graph$execute$3$1.invoke(Graph.kt)
at io.kinference.profiler.ProfilingContextKt.profile(ProfilingContext.kt:26)
at io.kinference.graph.Graph$execute$3.invokeSuspend(Graph.kt:223)
at io.kinference.graph.Graph$execute$3.invoke(Graph.kt)
at io.kinference.graph.Graph$execute$3.invoke(Graph.kt)
at kotlinx.coroutines.intrinsics.UndispatchedKt.startUndispatchedOrReturn(Undispatched.kt:78)
at kotlinx.coroutines.CoroutineScopeKt.coroutineScope(CoroutineScope.kt:264)
at io.kinference.graph.Graph.execute(Graph.kt:220)
at io.kinference.core.model.KIModel.predict(KIModel.kt:22)
at io.kinference.model.Model$DefaultImpls.predict$default(Model.kt:18)
...
at _COROUTINE.BOUNDARY.(CoroutineDebugging.kt:46)
at io.kinference.graph.Graph.execute(Graph.kt:220)
at io.kinference.core.model.KIModel.predict(KIModel.kt:22)

Could you take a look what is going wrong, please?

Would be glad to hear any feedback. Thanks!

[cupertank]

Hi, @codepictor! Could you send us this model file, please?

[codepictor]

Here it is.
rf_iris.zip

As I mentioned before, this file was produced by this: https://onnx.ai/sklearn-onnx/index.html.

[cupertank]

Thank you! We will take a look

[codepictor]

I would also like to emphasize that the model does not always fail. On some inputs it successfully produces the result, on other inputs it fails. For example:

FAIL:

val input = FloatNDArray(shape=intArrayOf(1, 4)) { (0.0).toFloat() }

val input = FloatNDArray(shape=intArrayOf(1, 4)) { (-1.0).toFloat() }

val input = FloatNDArray(shape=intArrayOf(1, 4)) { (it - 10).toFloat() }

SUCCESS:

val input = FloatNDArray(shape=intArrayOf(1, 4)) { (1.0).toFloat() }

val input = FloatNDArray(shape=intArrayOf(1, 4)) { (10.0).toFloat() }

val input = FloatNDArray(shape=intArrayOf(1, 4)) { (it - 2).toFloat() }

val input = FloatNDArray(shape=intArrayOf(1, 4)) { (it + 10).toFloat() }

[TimurReDoctor]

Hi @cupertank,
Do you have updates on that discussion?

[AnastasiaTuchina]

Hi! I'm working on it right now.

[AnastasiaTuchina]

Hi @TimurReDoctor!
We fixed this issue in 0.2.19 release

[codepictor]

Many thanks!

Could you, please, show how to initialize FloatNDArray properly or update your example regarding FloatNDArray? It seems that this example does not work anymore.

[AnastasiaTuchina]

You can use something like this
FloatNDArray(x, y) { index: InlineInt -> ... }

[codepictor]

I can confirm that now it works, and the produced results are the same as the results obtained in Python. Thanks a lot!