seq2seq

Sequence to Sequence Learning with Keras

Papers:

• Sequence to Sequence Learning with Neural Networks
• A Neural Conversational Model
• Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation

Notes:

• The LSTM Encoder encodes a sequence to a single a vector.

• The LSTM Decoder, when given a hidden state and a vector, generates a sequence.

• In the Seq2seq model, the output vector of the LSTM Encoder is the input for the LSTM Decoder, and

• The hidden state of the LSTM Encoder is copied to the hidden state of LSTM Decoder.

Continious VS Descrete sequence pairs:

• When training on continuous sequence pairs, such as long conversations, use the Conversational model instead of Seq2seq model, with argument context_sensitive=True. This is important if you want context sensitive conversational models, so that you can avoid scenarios like this:(Will only work if there are lot of exchanges in each conversation in your training data)

Human: what is your job ?

Machine: i ’m a lawyer .

Human: what do you do ?

Machine: i ’m a doctor

Source : A Neural Conversational Model

• When context_sensitive=True do not forget to clear the hidden state of Conversational layer after every conversation(Not after every exchange) or a fixed number of batches using reset_hidden_state() during training and testing. You could use the ResetState callback for this purpose.

• You will also have to clear the hidden state of Seq2seq layer after a fixed number of batches when used with remember_state=True.

• In case of descrete sequence pairs(for e.g, machine translation) use Seq2seq layer with the remeber_state argument set to False.

Example:

import keras
from keras.models import Sequential
from keras.layers.embeddings import Embedding
from seq2seq.seq2seq import Seq2seq
from keras.preprocessing import sequence

vocab_size = 20000 #number of words
maxlen = 100 #length of input sequence and output sequence
embedding_dim = 200 #word embedding size
hidden_dim = 500 #memory size of seq2seq

embedding = Embedding(vocab_size, embedding_dim, input_length=maxlen)
seq2seq = Seq2seq(input_length=maxlen, input_dim=embedding_dim,hidden_dim=hidden_dim,
                  output_dim=embedding_dim, output_length=maxlen, batch_size=10, depth=4)

model = Sequential()
model.add(embedding)
model.add(seq2seq)

Installation:

sudo pip install git+ssh://github.com/farizrahman4u/seq2seq.git

Requirements:

• Numpy
• Theano : Do not pip install
• Keras

Working Example:

• Training Seq2seq with movie subtitles - Thanks to Nicolas Ivanov