TF在馈入模型之前记录预处理数据

Question

我有一个包含此类功能的TFrecord文件

features = {
  'features': tf.FixedLenFeature([17920], tf.float32),
  'video_id': tf.FixedLenFeature([], tf.string),
  'split_id': tf.FixedLenFeature([], tf.int64),
  'audio': tf.FixedLenFeature([44100], tf.float32)
}

我目前获得迭代器的方式是这样的：

 def getiterator(batch_size, decode_parallel_calls=1,repeat=False,
    shuffle=False,
    shuffle_buffer_size=None,
    prefetch_size=None,
    prefetch_gpu_num=None):
dataset = tf.data.TFRecordDataset(['../main.tfrecords'])

  # Shuffle all filepaths every epoch
  if shuffle:
    dataset = dataset.shuffle(buffer_size=sum(1 for _ in tf.python_io.tf_record_iterator('../main.tfrecords')))

  # Repeat
  if repeat:
    dataset = dataset.repeat()

  # Shuffle examples
  if shuffle:
    dataset = dataset.shuffle(buffer_size=shuffle_buffer_size)

  def extract_fn(data_record):
    features = {
      'features': tf.FixedLenFeature([17920], tf.float32),
      'video_id': tf.FixedLenFeature([], tf.string),
      'split_id': tf.FixedLenFeature([], tf.int64),
      'audio': tf.FixedLenFeature([44100], tf.float32)
    }
    return tf.parse_single_example(data_record, features)

  dataset = dataset.map(extract_fn, num_parallel_calls=decode_parallel_calls)

  # Make batches
  dataset = dataset.batch(batch_size, drop_remainder=True)

  # Get tensors
  iterator = dataset.make_one_shot_iterator()

  return iterator.get_next()

我感兴趣的一件事是通过下采样到16000Hz来预处理音频数据。但是，我不确定该怎么做，因为我目前只获得张量。什么是好的方法？

Answer 1

您可以简单地将下采样逻辑放在extract_fn中。

def extract_fn(data_record):
    features = {
      'features': tf.FixedLenFeature([17920], tf.float32),
      'video_id': tf.FixedLenFeature([], tf.string),
      'split_id': tf.FixedLenFeature([], tf.int64),
      'audio': tf.FixedLenFeature([44100], tf.float32)
    }
    features_in = tf.parse_single_example(data_record, features)

    # This is your audio tensor before downsampling.
    audio_tensor = features_in['audio'] 
    # Doing something to downsample audio_tensor
    downsampled_audio_tensor = do_something(audio_tensor)

    # Now I assume you need every other tensors as is, but replace the audio with
    # the one we downsampled above.
    features_out = {
      'features': features_in['features'],
      'video_id': features_in['video_id'],
      'split_id': features_in['split_id'],
      'audio': downsampled_audio_tensor,
    }
    return features_out