video_input.py 文件源码

def __init__(self, config, data):
    self.batch_size = batch_size = config['batch_size']
    self.num_steps  = num_steps = config['num_steps']
    self.epoch_size = (data.num_examples_per_epoch() // batch_size) - 1
    # input_data size: [batch_size, num_steps]
    # targets size: [batch_size]
    self.input_data, self.targets, self.filenames = distorted_inputs(
      data, config)

    # Data preprocessing: input_data
    #  string tensor [batch_size, num_steps] =>
    #    num_steps * [batch_size, height*width*channels]
    self.input_data = tf.map_fn(
      decode_video, self.input_data, dtype=tf.float32)
    self.input_data = tf.reshape(
      self.input_data, [batch_size, num_steps, -1])
    self.input_data = [tf.squeeze(input_step, [1])
               for input_step in tf.split(self.input_data, num_steps, 1)]