multiscale_cnn_lstm_model.py 文件源码

def cnn(self, 
          model_input, 
          l2_penalty=1e-8, 
          num_filters = [1024, 1024, 1024],
          filter_sizes = [1,2,3], 
          sub_scope="",
          **unused_params):
    max_frames = model_input.get_shape().as_list()[1]
    num_features = model_input.get_shape().as_list()[2]

    shift_inputs = []
    for i in xrange(max(filter_sizes)):
      if i == 0:
        shift_inputs.append(model_input)
      else:
        shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])

    cnn_outputs = []
    for nf, fs in zip(num_filters, filter_sizes):
      sub_input = tf.concat(shift_inputs[:fs], axis=2)
      sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs, 
                       shape=[num_features*fs, nf], dtype=tf.float32, 
                       initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1), 
                       regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
      cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))

    cnn_output = tf.concat(cnn_outputs, axis=2)
    cnn_output = slim.batch_norm(
        cnn_output,
        center=True,
        scale=True,
        is_training=FLAGS.is_training,
        scope=sub_scope+"cluster_bn")
    return cnn_output