ValueError:GraphDef cannot be larger than 2GB.解决办法

在使用TensorFlow 1.X版本的estimator的时候经常会碰到类似于ValueError:GraphDef cannot be larger than 2GB的报错信息,可能的原因是数据太大无法写入graph。

一般来说,常见的数据构建方法如下:

def input_fn():
  features, labels = (np.random.sample((100,2)), np.random.sample((100,1)))
  dataset = tf.data.Dataset.from_tensor_slices((features,labels))
  dataset = dataset.shuffle(100000).repeat().batch(batch_size)
  return dataset

...
estimator.train(input_fn)

TensorFlow在读取数据的时候会将数据也写入Graph,所以当数据量很大的时候会碰到这种情况,之前做实验在多GPU的时候也会遇到这种情况,即使我把batch size调到很低。所以解决办法有两种思路,一直不保存graph,而是使用feed_dict的方式来构建input pipeline。

不写入graph

我的代码环境是TensorFlow1.14,所以我以这个版本为例进行介绍。

首先总结一下estimator的运行原理(假设在单卡情况下),以estimator.train为例(eval和predict类似),其调用顺序如下:

  1. estimator.train->_train_model

  2. _train_model->_train_model_default

  3. _train_model_default->_train_with_estimator_spec

  4. _train_with_estimator_spec->MonitoredTrainingSession

class Estimator():
	...
	def train():
		...
		loss = self._train_model(input_fn, hooks, saving_listeners)
		...
		
	def _train_model(self, input_fn, hooks, saving_listeners):
		if self._train_distribution:
			return self._train_model_distributed(input_fn, hooks, saving_listeners)
		else:
			return self._train_model_default(input_fn, hooks, saving_listeners)
	  
	def _train_model_default(self, input_fn, hooks, saving_listeners):
		...
		return self._train_with_estimator_spec(estimator_spec, worker_hooks,
											 hooks, global_step_tensor,
											 saving_listeners)
											 
	def _train_with_estimator_spec(self, estimator_spec, worker_hooks, hooks,
									 global_step_tensor, saving_listeners):
		....
		with training.MonitoredTrainingSession(
			master=self._config.master,
			is_chief=self._config.is_chief,
			checkpoint_dir=self._model_dir,
			scaffold=estimator_spec.scaffold,
			hooks=worker_hooks,
			chief_only_hooks=(tuple(chief_hooks) +
							  tuple(estimator_spec.training_chief_hooks)),
			save_checkpoint_secs=0,  # Saving is handled by a hook.
			save_summaries_steps=save_summary_steps,
			config=self._session_config,
			max_wait_secs=self._config.session_creation_timeout_secs,
			log_step_count_steps=log_step_count_steps) as mon_sess:

单步调试后发现,estimator写入event文件发生在调用MonitoredTrainingSession的时刻,而真正写入event是在执行hook的时候,例如在我的实验中我设置了log_step_count_steps这个值,这个值会每隔指定次数steps就会打印出计算速度和当前的loss值。而实现这一功能的是StepCounterHook,它定义在tensorflow/tensorflow/python/training/basic_session_run_hooks.py中,部分定义如下:

class StepCounterHook(session_run_hook.SessionRunHook):
  """Hook that counts steps per second."""

  def __init__(...):
  	...
    self._summary_writer = summary_writer
	
  def begin(self):
    if self._summary_writer is None and self._output_dir:
      self._summary_writer = SummaryWriterCache.get(self._output_dir)
    self._summary_tag = training_util.get_global_step().op.name + "/sec"

  def before_run(self, run_context):  # pylint: disable=unused-argument
    return SessionRunArgs(self._global_step_tensor)

  def _log_and_record(self, elapsed_steps, elapsed_time, global_step):
    steps_per_sec = elapsed_steps / elapsed_time
    if self._summary_writer is not None:
      summary = Summary(value=[
          Summary.Value(tag=self._summary_tag, simple_value=steps_per_sec)
      ])
      self._summary_writer.add_summary(summary, global_step)
    logging.info("%s: %g", self._summary_tag, steps_per_sec)

所以我们只需要将出现类似于self._summary_writer.add_summary的地方注释掉,这样estimator在运行过程中就不会再生成event文件,也就不会有2GB的问题了。

feed_dict

为了在大数据量时使用 dataset,我们可以用 placeholder 创建 dataset。这时数据就不会直接写到 graph 中,graph 中只有一个 placeholder 占位符。但是,用了 placeholder 就需要我们在一开始对它进行初始化填数据,需要调用 sess.run(iter.initializer, feed_dict={ x: data })

但是estimator并没有显示的session可以调用,那应该怎么办呢?其实我们可以使用SessionRunHook来解决这个问题。tf.train.SessionRunHook()类定义在tensorflow/python/training/session_run_hook.py,该类的具体介绍可参见【转】tf.SessionRunHook使用方法

仔细看一下 estimator 的 train 和 evaluate 函数定义可以发现它们都接收 hooks 参数,这个参数的定义是:List of tf.train.SessionRunHook subclass instances. Used for callbacks inside the training loop. 也就是说我们可以自己定义一个SessionRunHook作为参数传递到hook就可以了。

train(
    input_fn,
    hooks=None,
    steps=None,
    max_steps=None,
    saving_listeners=None
)

我们现在想要在训练之前初始化 dataset 的 placeholder,那么我们就应该具体实现 SessionRunHook 的after_create_session 成员函数:

class IteratorInitializerHook(tf.train.SessionRunHook):
   def __init__(self):
       super(IteratorInitializerHook, self).__init__()
       self.iterator_initializer_fn = None

   def after_create_session(self, session, coord):
       del coord
       self.iterator_initializer_fn(session)

def make_input_fn():
   iterator_initializer_hook = IteratorInitializerHook()

   def input_fn():
       x = tf.placeholder(tf.float32, shape=[None,2])
       dataset = tf.data.Dataset.from_tensor_slices(x)
       dataset = dataset.shuffle(100000).repeat().batch(batch_size)
       iter = dataset.make_initializable_iterator()
       data = np.random.sample((100,2))
       iterator_initializer_hook.iterator_initializer_fn = (
           lambda sess: sess.run(iter.initializer, feed_dict={x: data})
       )
       return iter.get_next()
   return input_fn, iterator_initializer_hook

...
input_fn, iterator_initializer_hook = make_input_fn()
estimator.train(input_fn, hooks=[iterator_initializer_hook])

参考



MARSGGBO♥原创





2019-10-21 11:04:22



上一篇:Generalized分位数回归, 新的前沿因果推断方法


下一篇:2-04-模型选择与调优