将有状态LSTM称为功能模型? [英] Calling a stateful LSTM as a functional model?

问题描述

我有一个定义为顺序模型的有状态LSTM:

I have a stateful LSTM defined as a Sequential model:

model = Sequential()
model.add(LSTM(..., stateful=True))
...

稍后，我将其用作功能模型:

Later, I use it as a Functional model:

input_1, input_2 = Input(...), Input(...)
output_1 = model(input_1)
output_2 = model(input_2)  # Is the state from input_1 preserved?

在input_2上再次应用model时，是否保留了input_1中的状态?如果是，如何在两次调用之间重置模型状态?

Is the state from input_1 preserved when we apply model again on input_2? If yes, how can I reset the model state in between the calls?

推荐答案

以下有关在RNN中使用状态的说明，来自此

Following Note on using statefulness in RNNs from this link and Keras implementation the answer is yes if:

1. 两个模型中的batch_size是相同的(由于Keras计算内部状态的方式，这一点很重要).
2. 您将首先构建和编译两个模型，然后使用它们-出于某种原因，Keras正在重置图层的build期间的内部状态(可以检查

1. The batch_size in both models is the same (it's important due to the way Keras computes the inner states).
2. You would first build and compile both models and then use them - for some reason Keras is resetting the inner states during the build of a layer (you can check it here by looking for reset_states method).

如果要重置状态，可以在每个要重置状态的循环层上调用reset_states方法.

If you want to reset states you could call reset_states method on each recurrent layer you want ot reset states on.

这篇关于将有状态LSTM称为功能模型?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！