How to build an RNN with input as character but output a sequence of tokens?

Question

Let’s assume I have “abbcab”-> {“Apple”,”Banana”}, “Bibb”-> {“Carrot”} so the output is a sequence of various length, but I can pad it to be the same length. My problem, how do I create a simple LSTM model that takes a string of characters of varying length and map it to a vector? What should be the last layer to return the vector?

Answer 1

This is a multilabel classification.

classes = Length@{"Apple", "Banana", "Carrot"};

net = NetChain[
  {
   EmbeddingLayer[8],
   LongShortTermMemoryLayer[16],
   SequenceLastLayer[],
   LinearLayer[classes],
   ElementwiseLayer[LogisticSigmoid]
   },
  "Input" -> NetEncoder[{"Characters"}]
  ];

netT = NetTrain[net, {"abbcab" -> {1, 1, 0}, "Bibb" -> {0, 0, 1}}, MaxTrainingRounds -> 100]

netT@"abbcab"

{0.822325, 0.864362, 0.163531}

netT@"Bibb"

{0.219002, 0.200206, 0.798121}

You should do your own post-processing to select top-n classes.