Tensorflow: 在方法中使用会话/图

Question

Tensorflow: 在方法中使用会话/图

5

我的情况是这样的：

我有一个训练tensorflow模型的脚本。在这个脚本中，我实例化了一个类来提供训练数据。这个类的初始化又会实例化另一个名为“image”的类，用于进行各种数据增强等操作。

main script -> instantiates data_feed class -> instantiates image class

我的问题是，我正在尝试使用tensorflow通过传递会话本身或图来在此图像类中执行一些操作。但我没有取得太大的成功。

可行的方法（但速度太慢）

我现在拥有的东西，但工作速度极慢，类似于这样（简化）：

class image(object):
    def __init__(self, im):
        self.im = im

    def augment(self):
        aux_im = tf.image.random_saturation(self.im, 0.6)

        sess = tf.Session(graph=aux_im.graph)
        self.im = sess.run(aux_im)

class data_feed(object):
    def __init__(self, data_dir):
        self.images = load_data(data_dir)

    def process_data(self):
        for im in self.images:
            image = image(im)
            image.augment()

if __name__ == "__main__":
    # initialize everything tensorflow related here, including model
    sess = tf.Session()
    # next load the data
    data_feed = data_feed(TRAIN_DATA_DIR)
    train_data = data_feed.process_data()

这种方法能够奏效，但它会为每个图像创建一个新的会话：

I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0)
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0)
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0)
I tensorflow/core/common_runtime/gpu/gpu_device.cc:975] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0)
etc ...

不起作用的方法（本应更快）

例如，以下做法并不起作用，我也无法弄清原因：从主脚本传递图或会话：

class image(object):
    def __init__(self, im):
        self.im = im

    def augment(self, tf_sess):
        with tf_sess.as_default():
            aux_im = tf.image.random_saturation(self.im, 0.6)

            self.im = tf_sess.run(aux_im)

class data_feed(object):
    def __init__(self, data_dir, tf_sess):
        self.images = load_data(data_dir)
        self.tf_sess = tf_sess

    def process_data(self):
        for im in self.images:
            image = image(im)
            image.augment(self.tf_sess)

if __name__ == "__main__":
    # initialize everything tensorflow related here, including model
    sess = tf.Session()
    # next load the data
    data_feed = data_feed(TRAIN_DATA_DIR, sess)
    train_data = data_feed.process_data()

这是我收到的错误信息：

Traceback (most recent call last):
  File "/usr/lib/python2.7/threading.py", line 801, in __bootstrap_inner
    self.run()
  File "/usr/lib/python2.7/threading.py", line 754, in run
    self.__target(*self.__args, **self.__kwargs)
  File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 409, in data_generator_task
    generator_output = next(generator)
  File "/home/mathetes/Dropbox/ML/load_gluc_data.py", line 198, in generate
    yield self.next_batch()
  File "/home/mathetes/Dropbox/ML/load_gluc_data.py", line 192, in next_batch
    X, y, l = self.process_image(json_im, X, y, l)
  File "/home/mathetes/Dropbox/ML/load_gluc_data.py", line 131, in process_image
    im.augment_with_tf(self.tf_sess)
  File "/home/mathetes/Dropbox/ML/load_gluc_data.py", line 85, in augment_with_tf
    self.im = sess.run(saturation, {im_placeholder: np.asarray(self.im)})
  File "/home/mathetes/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 766, in run
    run_metadata_ptr)
  File "/home/mathetes/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 921, in _run
    + e.args[0])
TypeError: Cannot interpret feed_dict key as Tensor: Tensor Tensor("Placeholder:0", shape=(96, 96, 3), dtype=float32) is not an element of this graph.

任何帮助都将不胜感激！

- mathetes

1个回答

网页内容由stack overflow 提供, 点击上面的

可以查看英文原文，
原文链接

- jabalazs · Accepted Answer

不如创建一个名为ImageAugmenter的类，它在初始化时需要一个会话，并使用Tensorflow处理图像，而不是使用Image类。你可以这样做：

import tensorflow as tf
import numpy as np

class ImageAugmenter(object):
    def __init__(self, sess):
        self.sess = sess
        self.im_placeholder = tf.placeholder(tf.float32, shape=[1,784,3])

    def augment(self, image):
        augment_op = tf.image.random_saturation(self.im_placeholder, 0.6, 0.8)
        return self.sess.run(augment_op, {self.im_placeholder: image})

class DataFeed(object):
    def __init__(self, data_dir, sess):
        self.images = load_data(data_dir)
        self.augmenter = ImageAugmenter(sess)

    def process_data(self):
        processed_images = []
        for im in self.images:
            processed_images.append(self.augmenter.augment(im))
        return processed_images

def load_data(data_dir):
    # True method would read images from disk
    # This is just a mockup
    images = []
    images.append(np.random.random([1,784,3]))
    images.append(np.random.random([1,784,3]))
    return images

if __name__ == "__main__":
    TRAIN_DATA_DIR = '/some/dir/'
    sess = tf.Session()
    data_feed = DataFeed(TRAIN_DATA_DIR, sess)
    train_data = data_feed.process_data()
    print(train_data)

使用这个方法，您不需要为每个图像创建一个新的会话，它应该能给您想要的结果。

请注意如何调用sess.run()；我传递给它的feed dict的键是上面定义的placeholder tensor。根据您的错误跟踪，您可能正在尝试从代码的某个部分调用sess.run()，而此时im_placeholder尚未定义，或者已经被定义为非tf.placeholder类型。

另外，您可以通过将下限和上限参数作为输入传递给tf.image.random_saturation()方法来进一步改进代码，或者您可以初始化ImageAugmenter以特定形状而不是硬编码，例如。