如何保存有状态的tensorflow keras模型的状态?

时间:2019-03-26 14:39:29

标签: python python-3.x tensorflow keras

我正在尝试编写一种新方法来分析时间序列预测的预测。因此,我需要在(预测的)每个时间戳复制我学习的模型,或者将其重置为上一个点,并用不同的输入进行输入。

我使用:

  • tensorflow.python.keras(tf版本:1.12)
  • python 3.7

我的网络:

model = Sequential()
model.add(GRU(100, return_sequences=True, stateful=True, batch_size=batchSize, input_shape=(x, y)))
model.add(Dropout(dropout))
model.add(GRU(units, return_sequences=False, stateful=True))
model.add(Dropout(dropout))
model.add(Dense(1, activation="linear", kernel_constraint=min_max_norm(min_value=-10)))
model.compile(loss='mean_squared_error', optimizer='nadam', metrics=['accuracy'])

目前唯一有效的方法是从头开始预测先前的步骤:

for i in range(timestamps):
    for j, features in enumerate(featuresTable[i]):
        if i > 0:
            model.predict(np.reshape(featuresList[:i], (i, 1, featuresList.shape[1])),
                                              batch_size=self.batch_size)
        predict = model.predict(np.reshape(features, (1, 1, len(features))), batch_size=self.batch_size)
        model.reset_states()                

其中timestamps是时间戳的数量,featuresTable是具有每个时间戳的替代功能的表,而featuresList是正常功能

我想要什么:

state = getState(model)
for i in range(timestamps):
    for j, features in enumerate(featuresTable[i]):
        predict = model.predict(np.reshape(features, (1, 1, len(features))), batch_size=self.batch_size)
        setState(model,state)
    model.predict(np.reshape(featuresList[i], (1, 1, featuresList.shape[1])), batch_size=self.batch_size)
    state = getState(model)           

先谢谢你 〜Lifree

0 个答案:

没有答案