python - 随着时间的推移，TensorFlow 中的图像转换速度变慢

Question

我有这种方法可以拍摄图像并将其转换为张量。我在一个循环中调用它，转换的执行时间开始很短并不断增长。

def read_tensor_from_image_file(file_name, input_height=299, input_width=299, input_mean=0, input_std=255):
    input_name = "file_reader"
    output_name = "normalized"
    file_reader = tf.read_file(file_name, input_name)

    image_reader = tf.image.decode_jpeg(file_reader, channels = 3,
        name='jpeg_reader')
    float_caster = tf.cast(image_reader, tf.float32)
    dims_expander = tf.expand_dims(float_caster, 0)
    resized = tf.image.resize_bilinear(dims_expander, [input_height, input_width])
    normalized = tf.divide(tf.subtract(resized, [input_mean]), [input_std])
    sess = tf.Session()
    result = sess.run(normalized)

    return result

我该如何优化它？

Answer 1

问题几乎可以肯定是由于在对 read_tensor_from_image_file() 函数的多次调用中使用了相同的默认 tf.Graph。解决此问题的最简单方法是在函数体周围添加一个 with tf.Graph().as_default(): block ，如下所示:

def read_tensor_from_image_file(file_name, input_height=299, input_width=299, input_mean=0, input_std=255):
   with tf.Graph().as_default():
    input_name = "file_reader"
    output_name = "normalized"
    file_reader = tf.read_file(file_name, input_name)
    image_reader = tf.image.decode_jpeg(file_reader, channels = 3,
        name='jpeg_reader')
    float_caster = tf.cast(image_reader, tf.float32)
    dims_expander = tf.expand_dims(float_caster, 0)
    resized = tf.image.resize_bilinear(dims_expander, [input_height, input_width])
    normalized = tf.divide(tf.subtract(resized, [input_mean]), [input_std])
    sess = tf.Session()
    result = sess.run(normalized)
    return result

通过此更改，每次调用该函数都会创建一个新图，而不是向默认图添加节点(默认图会随着时间增长，内存泄漏，并且每次使用它时需要更长时间才能启动)。

一个更高效的版本将使用 tf.placeholder() 作为文件名，构建单个图，并将 for 循环移动到 TensorFlow session 内部。像下面这样的东西会起作用:

def read_tensors_from_image_files(file_names, input_height=299, input_width=299, input_mean=0, input_std=255):
   with tf.Graph().as_default():
    input_name = "file_reader"
    output_name = "normalized"
    file_name_placeholder = tf.placeholder(tf.string, shape=[])
    file_reader = tf.read_file(file_name_placeholder, input_name)
    image_reader = tf.image.decode_jpeg(file_reader, channels = 3,
        name='jpeg_reader')
    float_caster = tf.cast(image_reader, tf.float32)
    dims_expander = tf.expand_dims(float_caster, 0)
    resized = tf.image.resize_bilinear(dims_expander, [input_height, input_width])
    normalized = tf.divide(tf.subtract(resized, [input_mean]), [input_std])

    with tf.Session() as sess:
      for file_name in file_names:
        yield sess.run(normalized, {file_name_placeholder: file_name})