Sequence-to-sequence models are composed of two separate components, an encoder and a decoder:
- Encoder: The encoder portion of the model takes an input sequence and returns an output and the network's internal state. We don't really care about the output; we only want to keep the encoder's state, which is the memory of the input sequence.
- Decoder: The decoder portion of the model then takes the state from the encoder, which is called the context or conditioning, as input. It then predicts the target sequence at each time step given the output of the previous time step.
The encoder and decoder then work together as pictured below, taking an input sequence and generating an output sequence. As you can see, we use special ...
