tensorflow - iterator.get_next() 导致在抛出 'std::system_error 实例后调用终止

Question

我正在使用 tensorflow 训练 resNet50，使用具有以下属性的共享服务器:

Ubuntu 16.04
3 gtx 1080 gpu
tensorflow 1.3
python 2.7
但总是在两个时代之后，在第三个时代，我遇到这个错误:

terminate called after throwing an instance of 'std::system_error' 
what():
Resource temporarily unavailable
Aborted 

这是将 tfrecord 转换为数据集的代码:

filenames = ["balanced_t.tfrecords"]
dataset = tf.contrib.data.TFRecordDataset(filenames)
def parser(record):
keys_to_features = {
    "mhot_label_raw": tf.FixedLenFeature((), tf.string, 
default_value=""),
    "mel_spec_raw": tf.FixedLenFeature((), tf.string, 
default_value=""),
}
parsed = tf.parse_single_example(record, keys_to_features)

mel_spec1d = tf.decode_raw(parsed['mel_spec_raw'], tf.float64)
# label = tf.cast(parsed["label"], tf.string)
mhot_label = tf.decode_raw(parsed['mhot_label_raw'], tf.float64)
mel_spec = tf.reshape(mel_spec1d, [96, 64])
return {"mel_data": mel_spec}, mhot_label
dataset = dataset.map(parser)
dataset = dataset.batch(batch_size)
dataset = dataset.repeat(3)
iterator = dataset.make_one_shot_iterator()

这是输入管道:

while True:
        try:
           (features, labels) = sess.run(iterator.get_next())
        except tf.errors.OutOfRangeError:
           print("end of training dataset")

在我的代码中插入一些打印消息后，我发现以下行会导致此错误:

(features, labels) = sess.run(iterator.get_next())

但是，我无法解决它

Answer 1

您的代码存在(微妙的)内存泄漏，因此该进程可能内存不足并被终止。问题是在每次循环迭代中调用 iterator.get_next() 都会向 TensorFlow 图中添加一个新节点，这最终会消耗大量内存。

要阻止内存泄漏，请按如下方式重写 while 循环:

# Call `get_next()` once outside the loop to create the TensorFlow operations once.
next_element = iterator.get_next()

while True:
    try:
        (features, labels) = sess.run(next_element)
    except tf.errors.OutOfRangeError:
        print("end of training dataset")