ModelSaver fails on saving Input with 2d and 3d tensors

[zaleslaw]

During saving model, trained on IrisData

const val SEED = 12L
const val TEST_BATCH_SIZE = 5
const val EPOCHS = 100
const val TRAINING_BATCH_SIZE = 3
const val PATH_TO_MODEL = "models/irisNet"

val model = Sequential.of(
    Input(4),
    Dense(5, Activations.Relu, kernelInitializer = HeNormal(SEED), biasInitializer = HeUniform(SEED)),
    Dense(3, Activations.Linear, kernelInitializer = HeNormal(SEED), biasInitializer = HeUniform(SEED))
)

fun main() {
    data.shuffle()

    val dataset = OnHeapDataset.create(
        ::extractX,
        ::extractY
    )

    val (train, test) = dataset.split(0.8)

    model.use {
        it.compile(optimizer = Adam(), loss = Losses.SOFT_MAX_CROSS_ENTROPY_WITH_LOGITS, metric = Metrics.ACCURACY)

        it.summary()
        it.fit(dataset = train, epochs = EPOCHS, batchSize = TRAINING_BATCH_SIZE)

        val accuracy = model.evaluate(dataset = test, batchSize = TEST_BATCH_SIZE).metrics[Metrics.ACCURACY]

        model.save(modelDirectory = File(PATH_TO_MODEL), savingFormat = SavingFormat.JSON_CONFIG_CUSTOM_VARIABLES, writingMode = WritingMode.OVERRIDE)

        println("Accuracy: $accuracy")
    }
}

it fails with the

Exception in thread "main" java.lang.IndexOutOfBoundsException: Index 1 out of bounds for length 1
	at java.base/jdk.internal.util.Preconditions.outOfBounds(Preconditions.java:64)
	at java.base/jdk.internal.util.Preconditions.outOfBoundsCheckIndex(Preconditions.java:70)
	at java.base/jdk.internal.util.Preconditions.checkIndex(Preconditions.java:248)
	at java.base/java.util.Objects.checkIndex(Objects.java:372)
	at java.base/java.util.ArrayList.get(ArrayList.java:459)
	at org.jetbrains.kotlinx.dl.api.inference.keras.ModelSaverKt.saveModelConfiguration(ModelSaver.kt:54)
	at org.jetbrains.kotlinx.dl.api.inference.keras.ModelSaverKt.saveModelConfiguration$default(ModelSaver.kt:36)
	at org.jetbrains.kotlinx.dl.api.core.GraphTrainableModel.saveModel(GraphTrainableModel.kt:794)
	at org.jetbrains.kotlinx.dl.api.core.GraphTrainableModel.saveInKerasFormat(GraphTrainableModel.kt:788)
	at org.jetbrains.kotlinx.dl.api.core.GraphTrainableModel.save(GraphTrainableModel.kt:783)
	at org.jetbrains.kotlinx.dl.api.core.TrainableModel.save$default(TrainableModel.kt:259)
	at IrisKt.main(Iris.kt:45)
	at IrisKt.main(Iris.kt)

due to unfixed TODO in ModelSaver

1. Please, fix the bug to handle not only images, but all situations to save the Input layer correctly to JSON config
2. It should work in both ways: for reading and for writing Input layer
3. Don't forget to add some tests for 1d and 3d data

[juliabeliaeva]

Hi, I'd like to work on this issue.

[zaleslaw]

Great, @juliabeliaeva feel free to ask any question in this issue or on Kotlin slack