In Sequence to Sequence is when a model takes sequence of information and outputs a sequence of information, for eg-language translation, image is converted into a text .
The two main components in Sequence to Sequence is Encoders and Decoders . Normally in encoders and decoders . They use these neural network like RNN,LSTM ,GRU in the encoder and decoder layer . In encoders the output will not be present and every word is given in each timestamp and at the last encoder its called the end of the string which gives an output which is a context vector .Each and every text is passed in the form of vectors which can be one hot representation, word embeddings,word2vec. Once t
he context vector is ready its passed to the decoder as an input and then we get the X as output and then the output will be passed as input to the next timestamp ,and then we get the Y as output and then the Y is passed as input to the Z and this continues and finally the end of string .The input sequence and the output sequence may be completely different in size .then we find the loss. we use the optimizer to reduce the loss using back propagation . The one disadvantage in encoder and decoder is that when the length of sentence increases the accuracy decreases (Blue score).
This is what encoder and decoder is ,and its also has few disadvantages to overcome that we can use attention models.