设置表示为 Keras 顺序模型的 RNN 的初始状态

Question

如何设置下面构建的循环神经网络rnn的初始状态？

from tensorflow.keras.layers import Dense, SimpleRNN
from tensorflow.keras.models import Sequential

rnn = Sequential([SimpleRNN(3), Dense(1)])

我想在用model.fit拟合模型之前指定第一层的初始状态。

Answer 1

根据tf.keras.layers.RNN documentation，您可以使用参数initial_state或数值符号指定初始状态] 通过调用函数 reset_states.

符号规范意味着您需要将初始状态添加为模型的输入。这是我改编自 Keras tests:

的示例

from tensorflow.keras.layers import Dense, SimpleRNN, Input
from tensorflow.keras.models import Model
import numpy as np
import tensorflow as tf

timesteps = 3
embedding_dim = 4
units = 3

inputs = Input((timesteps, embedding_dim))
# initial state as Keras Input
initial_state = Input((units,))
rnn = SimpleRNN(units)
hidden = rnn(inputs, initial_state=initial_state)
output = Dense(1)(hidden)

model = Model([inputs] + [initial_state], output)
model.compile(loss='categorical_crossentropy', 
              optimizer=tf.compat.v1.train.AdamOptimizer())

定义模型后，您可以按如下方式进行训练：

num_samples = 2

inputs = np.random.random((num_samples, timesteps, embedding_dim))
# random initial state as additional input
some_initial_state = np.random.random((num_samples, units))
targets = np.random.random((num_samples, units))
model.train_on_batch([inputs] + [some_initial_state], targets)

请注意，此方法要求您使用 Functional API。对于顺序模型，您需要使用 stateful RNN，指定 batch_input_shape，并调用 reset_states 方法：

input_shape = (num_samples, timesteps, embedding_dim)
model = Sequential([
    SimpleRNN(3, stateful=True, batch_input_shape=input_shape), 
    Dense(1)])

some_initial_state = np.random.random((num_samples, units))
rnn = model.layers[0]
rnn.reset_states(states=some_initial_state)

设置表示为 Keras 顺序模型的 RNN 的初始状态

Setting the initial state of an RNN represented as a Keras sequential model

python

neural-network

keras

tensorflow

recurrent-neural-network