LSTM - 2. decode_training_set

# Decoding the training set
def decode_training_set(encoder_state, decoder_cell, decoder_embedded_input, sequence_length, decoding_scope, output_function, keep_prob, batch_size):
    attention_states = tf.zeros([batch_size, 1, decoder_cell.output_size])                                                
    attention_keys, attention_values, attention_score_function, attention_construct_function = tf.contrib.seq2seq.prepare_attention(init_attention_states, attention_option = "bahdanau", num_units = decoder_cell.output_size)
    training_decoder_function = tf.contrib.seq2seq.attention_decoder_fn_train(encoder_state[0],
                                                                              attention_keys,
                                                                              attention_values,
                                                                              attention_score_function,
                                                                              attention_construct_function,
                                                                              name = "attn_dec_train")
    decoder_output, decoder_final_state, decoder_final_context_state = tf.contrib.seq2seq.dynamic_rnn_decoder(decoder_cell,
                                                                                                              training_decoder_function,
                                                                                                              decoder_embedded_input,
                                                                                                              sequence_length,
                                                                                                              scope = decoding_scope)
    decoder_output_dropout = tf.nn.dropout(decoder_output, keep_prob)
    return output_function(decoder_output_dropout)

init_attention_states = tf.zeros([batch_size, 1, num_hidden_neurons])

Here, we set num_hidden_neurons as decoder_cell.output_size.

attention_keys, attention_values, attention_score_function, attention_construct_function = 
    tf.contrib.seq2seq.prepare_attention(
        init_attention_states, # init hidden states
        attention_option = "bahdanau", 
        num_units = decoder_cell.output_size # hidden state dimension
    )