multiscale_cnn_lstm_model.py 文件源码

def create_model(self, model_input, vocab_size, num_frames, 
                   l2_penalty=1e-8, **unused_params):

    num_layers = FLAGS.multiscale_cnn_lstm_layers
    lstm_size = int(FLAGS.lstm_cells)
    pool_size=2
    num_filters=[256,256,512]
    filter_sizes=[1,2,3]
    features_size = sum(num_filters)

    sub_predictions = []
    cnn_input = model_input

    cnn_max_frames = model_input.get_shape().as_list()[1]

    for layer in range(num_layers):
      cnn_output = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
      cnn_output_relu = tf.nn.relu(cnn_output)

      lstm_memory = self.rnn(cnn_output_relu, lstm_size, num_frames, sub_scope="rnn%d"%(layer+1))
      sub_prediction = self.moe(lstm_memory, vocab_size, scopename="moe%d"%(layer+1))
      sub_predictions.append(sub_prediction)

      cnn_max_frames /= pool_size
      max_pooled_cnn_output = tf.reduce_max(
          tf.reshape(
              cnn_output_relu[:, :cnn_max_frames*2, :], 
              [-1, cnn_max_frames, pool_size, features_size]
          ), axis=2)

      # for the next cnn layer
      cnn_input = max_pooled_cnn_output
      num_frames = tf.maximum(num_frames/pool_size, 1)

    support_predictions = tf.concat(sub_predictions, axis=1)
    predictions = tf.add_n(sub_predictions) / len(sub_predictions)

    return {"predictions": predictions, 
            "support_predictions": support_predictions}