键盘中断 tensorflow 运行并在那时保存

Question

有没有办法通过键盘中断来中断 tensorflow 会话并可以选择在那时保存模型？我目前让会话在一夜之间运行，但需要停止它以便我可以在白天释放内存以供 PC 使用。随着训练的进行，每个 epoch 都会变慢，因此有时我可能需要等待数小时才能在程序中进行下一次预定保存。我想要能够随时进入运行并从那时起保存的功能。我什至找不到这是否可能。将不胜感激指针。

Answer 1

一种选择是将tf.Session对象子类化并创建一个__exit__函数，该函数在键盘中断通过时保存当前状态。这仅在新对象作为with块的一部分被调用时才有效。

这是子类：

import tensorflow as tf

class SessionWithExitSave(tf.Session):
    def __init__(self, *args, saver=None, exit_save_path=None, **kwargs):
        self.saver = saver
        self.exit_save_path = exit_save_path
        super().__init__(*args, **kwargs)

    def __exit__(self, exc_type, exc_value, exc_tb):
        if exc_type is KeyboardInterrupt:
            if self.saver:
                self.saver.save(self, self.exit_save_path)
                print('Output saved to: "{}./*"'.format(self.exit_save_path))
        super().__exit__(exc_type, exc_value, exc_tb)

TensorFlow mnist 演练中的示例用法。

import tensorflow as tf
import datetime as dt
from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets('U:/mnist/', one_hot=True)
x = tf.placeholder(tf.float32, [None, 784])
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
y = tf.matmul(x, W) + b
# Define loss and optimizer
y_ = tf.placeholder(tf.float32, [None, 10])
cross_entropy = tf.reduce_mean(
    tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))
train_step = tf.train.GradientDescentOptimizer(0.2).minimize(cross_entropy)

saver = tf.train.Saver()

with SessionWithExitSave(
        saver=saver, 
        exit_save_path='./tf-saves/_lastest.ckpt') as sess:
    sess.run(tf.global_variables_initializer())
    total_epochs = 50
    for epoch in range(1, total_epochs+1):
        for _ in range(1000):
            batch_xs, batch_ys = mnist.train.next_batch(100)
            sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
        # Test trained model
        correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

        print(f'Epoch {epoch} of {total_epochs} :: accuracy = ', end='')
        print(sess.run(accuracy, feed_dict={x: mnist.test.images, y_: mnist.test.labels}))
        save_time = dt.datetime.now().strftime('%Y%m%d-%H.%M.%S')
        saver.save(sess, f'./tf-saves/mnist-{save_time}.ckpt')

在从键盘发送中断信号之前，我让它运行了 10 个 epoch。这是输出：

Epoch 1 of 50 :: accuracy = 0.9169
Epoch 2 of 50 :: accuracy = 0.919
Epoch 3 of 50 :: accuracy = 0.9205
Epoch 4 of 50 :: accuracy = 0.9221
Epoch 5 of 50 :: accuracy = 0.92
Epoch 6 of 50 :: accuracy = 0.9229
Epoch 7 of 50 :: accuracy = 0.9234
Epoch 8 of 50 :: accuracy = 0.9234
Epoch 9 of 50 :: accuracy = 0.9252
Epoch 10 of 50 :: accuracy = 0.9248
Output saved to: "./tf-saves/_lastest.ckpt./*"
---------------------------------------------------------------------------
KeyboardInterrupt                         Traceback (most recent call last)
...
--> 768   elif item[0].cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
    769     return item[1]._is_present_in_parent
    770   else:
KeyboardInterrupt:

事实上，我确实拥有所有保存的文件，包括从发送到系统的键盘中断保存的文件。

import os

os.listdir('./tf-saves/')
# returns:
['checkpoint',
 'mnist-20171207-23.05.18.ckpt.data-00000-of-00001',
 'mnist-20171207-23.05.18.ckpt.index',
 'mnist-20171207-23.05.18.ckpt.meta',
 'mnist-20171207-23.05.22.ckpt.data-00000-of-00001',
 'mnist-20171207-23.05.22.ckpt.index',
 'mnist-20171207-23.05.22.ckpt.meta',
 'mnist-20171207-23.05.26.ckpt.data-00000-of-00001',
 'mnist-20171207-23.05.26.ckpt.index',
 'mnist-20171207-23.05.26.ckpt.meta',
 'mnist-20171207-23.05.31.ckpt.data-00000-of-00001',
 'mnist-20171207-23.05.31.ckpt.index',
 '_lastest.ckpt.data-00000-of-00001',
 '_lastest.ckpt.index',
 '_lastest.ckpt.meta']