GRU for timeseries forcasting

[charlie1986]

I am trying to translate a python code using tensorflow/keras to DJL equivalent code. Below is python code:
`# Architecture Gated Recurrent Unit
regressorGRU = Sequential()

//First GRU layer with dropout
regressorGRU.add(GRU(units=hidden_unit, return_sequences=True, input_shape=(x_train.shape[1],1), activation = 'tanh'))
regressorGRU.add(Dropout(0.2))
//Second GRU layer with dropout
regressorGRU.add(GRU(units=hidden_unit, return_sequences=True, activation = 'tanh'))
regressorGRU.add(Dropout(0.2))
//Third GRU layer with dropout
regressorGRU.add(GRU(units=hidden_unit, return_sequences=False, activation = 'tanh'))
regressorGRU.add(Dropout(0.2))

//Output layer
regressorGRU.add(Dense(units=1))`

And here the DJL code:
public static SequentialBlock createNeuralNetwork(int hiddenUnit) {
//Architecture Gated Recurrent Unit
var regressorGRU = new SequentialBlock();
//First GRU layer with dropout
var gru1 = GRU.builder()
.setNumLayers(1)
.setStateSize(hiddenUnit)
.optReturnState(true)
.build();
regressorGRU.add(gru1);
regressorGRU.add(Activation::tanh);
regressorGRU.add(Dropout.builder().optRate(0.2f).build());
//Second GRU layer with dropout
var gru2 = GRU.builder()
.setNumLayers(1)
.setStateSize(hiddenUnit)
.optReturnState(true)
.build();
regressorGRU.add(gru2);
regressorGRU.add(Activation::tanh);
regressorGRU.add(Dropout.builder().optRate(0.2f).build());
//Third GRU layer with dropout
var gru3 = GRU.builder()
.setNumLayers(1)
.setStateSize(hiddenUnit)
.optReturnState(false)
.build();
regressorGRU.add(gru3);
regressorGRU.add(Activation::tanh);
regressorGRU.add(Dropout.builder().optRate(0.2f).build());
//Output layer
regressorGRU.add(Linear.builder().setUnits(1).build());
return regressorGRU;
}
But the error has occured: java.lang.UnsupportedOperationException: Expected layout: NTC, but got: ??
How to implement the equivalent code in DJL?

[zachgk]

This error is saying that the recurrent block expects an input of Shape(batchSize, timeSteps, channels), but it received 2-dimensional data. Is the input you are passing it correct? You may need to reshape the input so it has a 1 for one of those dimensions