如何根据当前纪元或纪元步骤编写某些层的逻辑? [英] How can I write some layer's logic based on the current epoch or step of epoch?
问题描述
Keras的文档在此处中介绍了如何编写自定义层，方法是从Layer类继承.现在，我有另一个自定义图层CustomLayer，我想从中继承该图层.我们将其称为新的自定义图层CustomLayer2.我猜想，即使我不是从Layer继承而是从CustomLayer继承，继承过程也将与链接文档中描述的相同，但是我还将继承CustomLayer的自定义功能.
The Keras' documentation describes here how to write a custom layer by inheriting from the Layer class. Now, I have another custom layer CustomLayer, from which I want to inherit. Let's call my new custom layer CustomLayer2. I guess that the process of inheriting, even though I will not be inheriting from Layer but from CustomLayer, will be the same described in the linked documentation, but I will also inherit the custom functionality of CustomLayer.
无论如何，在这个自定义层CustomLayer2内，我想实现一些基于当前纪元或纪元步骤的逻辑.
Anyway, inside this custom layer CustomLayer2 I want to implement some logic that is based on the current epoch or step of the epoch.
我该怎么做?如何从图层的call和/或build方法动态获取当前纪元或纪元的步骤?
How can I do that? How can I dynamically get the current epoch or step of the epoch from the call and/or build methods of the layer?
也许可以通过回调来完成.例如，我可以有一个回调，该回调访问模型的实例，然后在模型内部进行更改.我不太喜欢这种解决方案，但是如果可行，那就很好.但是我们可以从模型实例更改模型各层的逻辑吗?
Maybe this can be done with callbacks. For example, I could have a callback that accesses an instance of the model and then changes something inside the model. I don't like much this solution, but if it works, that's fine. But can we change the logic of the layers of a model from an instance of a model?
过去，有人问过两个类似的问题
In the past, two similar questions have been asked
- Tensorflow Keras modify model variable from callback
- Can I access what was once `tf.get_global_step()` from within a custom Keras layer?
推荐答案
自定义回调将是前往此处的方法.通常，层逻辑被调用一次以构造计算功能逻辑.稍后将在训练期间使用此计算功能.同样，这通常是在优化程序开始运行之前发生的，因此与时代/步骤无关的事物都没有.
A custom callback would be the way to go here. Generally, layer logic is called once to construct computation function logic; this computational function will later be used during training. This, again, generally, happens before optimizer comes to life and therefore nothing related to epochs/steps is around.
因此，无法从自定义图层中获取训练步骤.
So, there is no way to get training steps from inside a custom layer.
这篇关于如何根据当前纪元或纪元步骤编写某些层的逻辑?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！
