multires_lstm_memory_deep_combine_chain_model.py 文件源码

def lstm(self, model_input, vocab_size, num_frames, sub_scope="", **unused_params):
    number_of_layers = FLAGS.lstm_layers
    lstm_sizes = map(int, FLAGS.lstm_cells.split(","))
    feature_names, feature_sizes = utils.GetListOfFeatureNamesAndSizes(
        FLAGS.feature_names, FLAGS.feature_sizes)
    sub_inputs = [tf.nn.l2_normalize(x, dim=2) for x in tf.split(model_input, feature_sizes, axis = 2)]

    assert len(lstm_sizes) == len(feature_sizes), \
      "length of lstm_sizes (={}) != length of feature_sizes (={})".format( \
      len(lstm_sizes), len(feature_sizes))

    states = []
    for i in xrange(len(feature_sizes)):
      with tf.variable_scope(sub_scope+"RNN%d" % i):
        sub_input = sub_inputs[i]
        lstm_size = lstm_sizes[i]
        ## Batch normalize the input
        stacked_lstm = tf.contrib.rnn.MultiRNNCell(
                [
                    tf.contrib.rnn.BasicLSTMCell(
                        lstm_size, forget_bias=1.0, state_is_tuple=True)
                    for _ in range(number_of_layers)
                    ],
                state_is_tuple=True)
        output, state = tf.nn.dynamic_rnn(stacked_lstm, sub_input,
                                         sequence_length=num_frames,
                                         swap_memory=FLAGS.rnn_swap_memory,
                                         dtype=tf.float32)
        states.extend(map(lambda x: x.c, state))
    final_state = tf.concat(states, axis = 1)
    return final_state