vgg_ucf_preprocessing.py 文件源码

def preprocess_for_train(image,
                         output_height,
                         output_width,
                         mean_vals,
                         out_dim_scale=1.0):
  """Preprocesses the given image for training.

  Note that the actual resizing scale is sampled from
    [`resize_size_min`, `resize_size_max`].

  Args:
    image: A `Tensor` representing an image of arbitrary size.
    output_height: The height of the image after preprocessing.
    output_width: The width of the image after preprocessing.

  Returns:
    A preprocessed image.
  """
  num_channels = image.get_shape().as_list()[-1]
  image = tf.image.resize_images(image, [_RESIZE_HT, _RESIZE_WD])
  # compute the crop size
  base_size = float(min(_RESIZE_HT, _RESIZE_WD))
  scale_ratio_h = tf.random_shuffle(tf.constant(_SCALE_RATIOS))[0]
  scale_ratio_w = tf.random_shuffle(tf.constant(_SCALE_RATIOS))[0]
  image = _random_crop([image],
      tf.cast(output_height * scale_ratio_h, tf.int32),
      tf.cast(output_width * scale_ratio_w, tf.int32))[0]
  image = tf.image.resize_images(
    image, [int(output_height * out_dim_scale),
            int(output_width * out_dim_scale)])
  image = tf.to_float(image)
  image = tf.image.random_flip_left_right(image)
  image.set_shape([int(output_height * out_dim_scale),
                   int(output_width * out_dim_scale), num_channels])
  image = _mean_image_subtraction(image, mean_vals)
  image = tf.expand_dims(image, 0) # 1x... image, to be consistent with eval
  # Gets logged multiple times with NetVLAD, so gives an error.
  # I'm anyway logging from the train code, so removing it here.
  # tf.image_summary('final_distorted_image',
  #     tf.expand_dims(image / 128.0, 0))
  return image