decoders

Type Members

1. class BasicDecoder[O, OS, S, SS] extends Decoder[O, OS, S, SS, decoders.BasicDecoder.Output[O, OS], (OS, OS), S, SS, decoders.BasicDecoder.Output[O, OS], S]

   

   Basic sampling Recurrent Neural Network (RNN) decoder.

2. class BeamSearchDecoder[S, SS] extends Decoder[Output, core.Shape, S, SS, decoders.BeamSearchDecoder.Output, (core.Shape, core.Shape, core.Shape), State[S, SS], (SS, core.Shape, core.Shape, core.Shape), FinalOutput, State[S, SS]]

   

   Recurrent Neural Network (RNN) that uses beam search to find the highest scoring sequence (i.e., perform decoding).

3. abstract class Decoder[O, OS, S, SS, DO, DOS, DS, DSS, DFO, DFS] extends AnyRef

   

   Recurrent Neural Network (RNN) decoder abstract interface.

   Recurrent Neural Network (RNN) decoder abstract interface.

   Concepts used by this interface:

   • input: (structure of) tensors and tensor arrays that is passed as input to the RNN cell composing the decoder, at each time step.
   • state: Sequence of tensors that is passed to the RNN cell instance as the state.
   • finished: Boolean tensor indicating whether each sequence in the batch has finished decoding.

4. case class ExponentialPenalty(alpha: Float) extends LengthPenalty with Product with Serializable

   

   Exponential length penalty function.

   Exponential length penalty function. The penalty is equal to sequenceLengths ^ alpha, where all operations a re performed element-wise.

   alpha

   Length penalty weight (disabled if set to 0.0f).

5. case class GooglePenalty(alpha: Float) extends LengthPenalty with Product with Serializable

   

   Google length penalty function described in [Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation](https://arxiv.org/abs/1609.08144.) The penalty is equal to ((5 + sequenceLengths) / 6) ^ alpha, where all operations are performed element-wise.

   Google length penalty function described in [Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation](https://arxiv.org/abs/1609.08144.) The penalty is equal to ((5 + sequenceLengths) / 6) ^ alpha, where all operations are performed element-wise.

   alpha

   Length penalty weight (disabled if set to 0.0f).

6. trait LengthPenalty extends AnyRef

   

   Length penalty function to be used while decoding.