如何为 LSTM keras 重塑 X_train 和 y_train

Question

我有以下内容：

X_train.shape
(2730, 10)

y_train.shape
(2730)

我想用 keras 训练 LSTM 模型，但我不确定如何重塑输入。

我添加了这个 LSTM 层

time_steps = 30
input_dim = 10 # number of features
...
self.model.add(LSTM(self.hidden_dim, input_shape=(time_steps, self.input_dim), return_sequences=True))
...

input_shape 与我的输入不符。我应该如何重塑我的 X_train？我还必须重塑 y_train 吗？

Answer 1

How should I reshape my X_train?

最简单的选择是向数据添加 timesteps 维度，使其与 LSTM:

兼容

import tensorflow as tf

samples = 5
features = 10
data = tf.random.normal((samples, features))
time_series_data = tf.expand_dims(data, axis=1) # add timesteps dimension
tf.print('Data -->', tf.shape(data), 'Time series data', tf.shape(time_series_data))
# Data --> [5 10] Time series data [5 1 10]

但是，您希望对通向形状 (samples, 30, 10) 的每个特征使用 30 timesteps。所以，你可以使用的是RepeatVector layer as part of your model or tf.repeat。这是 RepeatVector 层的示例：

model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(10, input_shape=(features,)))
model.add(tf.keras.layers.RepeatVector(30))
model.add(tf.keras.layers.LSTM(32))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))
model.build((1, 10))
tf.print(model.summary())

Model: "sequential_01"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_24 (Dense)            (None, 10)                110       
                                                                 
 repeat_vector_1 (RepeatVect  (None, 30, 10)           0         
 or)                                                             
                                                                 
 lstm_3 (LSTM)               (None, 32)                5504      
                                                                 
 dense_25 (Dense)            (None, 1)                 33        
                                                                 
=================================================================
Total params: 5,647
Trainable params: 5,647
Non-trainable params: 0
_________________________________________________________________
None

您也可以先将 10 个特征映射到 300 维输出，然后重塑输出以适应 LSTM:

model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(300, input_shape=(features,)))
model.add(tf.keras.layers.Reshape((30, 10)))
model.add(tf.keras.layers.LSTM(32))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))

Model: "sequential_02"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense_26 (Dense)            (None, 300)               3300      
                                                                 
 reshape (Reshape)           (None, 30, 10)            0         
                                                                 
 lstm_4 (LSTM)               (None, 32)                5504      
                                                                 
 dense_27 (Dense)            (None, 1)                 33        
                                                                 
=================================================================
Total params: 8,837
Trainable params: 8,837
Non-trainable params: 0
_________________________________________________________________
None

对问题：

Do I also have to reshape the y_train?

这取决于你想要什么。如果你只有一个简单的分类任务，就像我在示例中假设的那样，那么你不需要更改 y_train.

更新一： 您还可以像下面这样重塑数据。产生具有 91 个样本的张量，其中每个样本有 30 个时间步长，每个时间步长与 10 个特征相关联。

import tensorflow as tf

timesteps = 2730
features = 10
data = tf.random.normal((timesteps, features))
data = tf.reshape(data, (91, 30, features))
print(data.shape)
# (91, 30, 10)

如何为 LSTM keras 重塑 X_train 和 y_train

How to reshape X_train and y_train for LSTM keras

python

lstm

keras

tensorflow