Seq2Seq model

Seq2Seq, the full name is Sequence to Sequence model,the meaning can be interpreted as a sequence signal(long sentences, paragraphs, image extraction features, audio signals,etc.),through encoding and decoding to generate a new sequence signal(abbreviated phrase, text description, recognition text), usually used in the fields of machine translation, picture description, automatic dialogue, speech recognition, etc.

What is Seq2Seq model?

The core idea of the seq2Seq model is to convert a sequence signal as an input to an output sequence signal through a deep neural network. This process consists of two processes:

Encoding and Decoding. In the classic implementation, the encoder and decoder are each composed of a recurrent neural network(RNN, LSTM, GRU can be).In Seq2Seq, the two recurrent neural networks are trained together.

The following figure describes several Seq2Seq models:

one to one: One input predicts one output(such as a picture to predict picture classification), one to one network structure is as follows:

one to many: One input predicts a sequence output (such as: (Picture description, input a picture to predict a text description of a sequence of this picture) One to many network structure is as follows:

(1)We only enter the input once. (2)We entered unique input once for each hidden layer

many to one: A sequence input gets an output (such as sentiment analysis, input a sequence of text of a length to get the evaluation attitude of the text is positive or negative)

many to many: A sequence input predicts a sequence output (such as (Machine translation, automatic dialogue generation)