Deep Learning Quick Reference

Decoding a sequence

The decoder needs the following two things to start its work:

• State from the encoder.
• An input signal that starts the predicted translation. We will send it a '\t' in a one hot vector as that's our <SOS> character.

To obtain the encoder state, we just need to send the vectorized version of the phrase we want to translate to the encoder, using the following code:

states_value = encoder_model.predict(input_seq)

In order to start the decoder, we also need a one hot vector containing the <SOS> character. This code gets us there:

target_seq = np.zeros((1, 1, data['num_decoder_tokens']))target_seq[0, 0, data['target_token_index']['\t']] = 1.

And now we're ready to set up a decoder loop that will generate our translated phrase, ...