使用估算器时用 tf.train.SessionRunHook 替换验证监视器 [英] Replace Validation Monitors with tf.train.SessionRunHook when using Estimators

问题描述

我正在运行 DNNClassifier，我在训练时对其准确性进行监控.contrib/learn 中的 monitor.ValidationMonitor 运行良好，在我的实现中我定义了它:

I am running a DNNClassifier, for which I am monitoring accuracy while training. monitors.ValidationMonitor from contrib/learn has been working great, in my implementation I define it:

validation_monitor = skflow.monitors.ValidationMonitor(input_fn=lambda: input_fn(A_test, Cl2_test), eval_steps=1, every_n_steps=50)

然后使用调用来自:

clf.fit(input_fn=lambda: input_fn(A, Cl2),
            steps=1000, monitors=[validation_monitor])

哪里:

clf = tensorflow.contrib.learn.DNNClassifier(...

这很好用.也就是说，验证监视器似乎已被弃用，类似的功能将被替换为 tf.train.SessionRunHook.

This works fine. That said, validation monitors appear to be deprecated and a similar functionality to be replaced with tf.train.SessionRunHook.

我是 TensorFlow 的新手，对我来说，这种替换实现的样子似乎并不简单.任何建议都受到高度赞赏.同样，我需要在特定数量的步骤后验证训练.非常感谢.

I am a newbie in TensorFlow, and it does not seem trivial to me how such a replacing implementation would look like. Any suggestion are highly appreciated. Again, I need to validate the training after a specific number of steps. Thanks very much in advance.

推荐答案

有一个未公开的实用工具 monitors.replace_monitors_with_hooks() 将监视器转换为钩子.该方法接受 (i) 一个可能包含监视器和钩子的列表，以及 (ii) 将使用钩子的 Estimator，然后通过在每个监视器周围包装一个 SessionRunHook 来返回一个钩子列表.

There's an undocumented utility called monitors.replace_monitors_with_hooks() which converts monitors to hooks. The method accepts (i) a list which may contain both monitors and hooks and (ii) the Estimator for which the hooks will be used, and then returns a list of hooks by wrapping a SessionRunHook around each Monitor.

from tensorflow.contrib.learn.python.learn import monitors as monitor_lib

clf = tf.estimator.Estimator(...)

list_of_monitors_and_hooks = [tf.contrib.learn.monitors.ValidationMonitor(...)]
hooks = monitor_lib.replace_monitors_with_hooks(list_of_monitors_and_hooks, clf)

对于完全替换 ValidationMonitor 的问题，这并不是真正的真正解决方案——我们只是用一个未弃用的函数将其包装起来.但是，我可以说到目前为止这对我有用，因为它保留了我从 ValidationMonitor 中需要的所有功能(即评估每个 n 步骤，使用指标提前停止等)

This isn't really a true solution to the problem of fully replacing the ValidationMonitor—we're just wrapping it up with a non-deprecated function instead. However, I can say this has worked for me so far in that it maintained all the functionality I need from the ValidationMonitor (i.e. evaluating every n steps, early stopping using a metric, etc.)

还有一件事——要使用这个钩子，你需要从 tf.contrib.learn.Estimator(只接受监视器)更新到更成熟和官方的 tf.estimator.Estimator(只接受钩子).因此，您应该将分类器实例化为 tf.estimator.DNNClassifier，并使用其方法 train() 进行训练(这只是 的重命名)fit()):

One more thing—to use this hook you'll need to update from a tf.contrib.learn.Estimator (which only accepts monitors) to the more full-fledged and official tf.estimator.Estimator (which only accepts hooks). So, you should instantiate your classifier as a tf.estimator.DNNClassifier, and train using its method train() instead (which is just a re-naming of fit()):

clf = tf.estimator.Estimator(...)

...

clf.train(
    input_fn=...
    ...
    hooks=hooks)

这篇关于使用估算器时用 tf.train.SessionRunHook 替换验证监视器的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！