How to realize realtime classification for bidirectional lstm?

layers = [
    sequenceInputLayer(64, Minlength=4)
    bilstmLayer(20)
    bilstmLayer(10)
    fullyConnectedLayer(2)
    softmaxLayer
    classificationLayer(2)]

I have trained a model using `layers`, everything work fine while doing forward inferencing using `classify`, when change to `classifyAndUpdateState` got wrong classification result.

My question is how to do real time frame by frame classification for bidirectional lstm? More specifically convert to C code using codegen.