如何为 LSTM 层的循环神经网络重塑张量

Question

我正在尝试训练 RNN，我的 X 输入形状是 (5018, 481)，而 y 标签输入形状是 (5018,)。我已将 X 和 y 都转换为以下格式的张量：

x_train_tensor = tf.convert_to_tensor(X, dtype=tf.float32)
y_train_tensor = tf.convert_to_tensor(y, dtype=tf.float32)

然后用下面的RNN keras模型架构：

model = keras.Sequential([
    keras.layers.Dense(100, activation='elu', input_shape=(481,)),
    keras.layers.LSTM(64, return_sequences=False, dropout=0.1, recurrent_dropout=0.1),
    keras.layers.Dense(25, activation='elu'),
    keras.layers.Dropout(0.5),
    keras.layers.Dense(1, 'elu')
])

opt = keras.optimizers.Adam(lr=0.001)
model.compile(optimizer=opt, loss='mean_squared_error', metrics=['mse'])

model.fit(x_train_tensor, y_train_tensor, epochs=8)

我收到以下错误

ValueError: Input 0 of layer lstm_1 is incompatible with the layer: 
expected ndim=3, found ndim=2. Full shape received: [None, 100]

有人有解决办法吗？

Answer 1

它期望得到 3 个维度：如果您打印没有 LSTM 层的网络摘要：

model = keras.Sequential([
keras.layers.Dense(100, activation='elu', input_shape=(481,)),
keras.layers.Dense(25, activation='elu'),
keras.layers.Dropout(0.5),
keras.layers.Dense(1,activation= 'elu')
])

opt = keras.optimizers.Adam(lr=0.001)
model.compile(optimizer=opt, loss='mean_squared_error', metrics=['mse'])
model.summary()

你得到：

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_19 (Dense)             (None, 100)               48200     
_________________________________________________________________
dense_20 (Dense)             (None, 25)                2525      
_________________________________________________________________
dropout_7 (Dropout)          (None, 25)                0         
_________________________________________________________________
dense_21 (Dense)             (None, 1)                 26        
=================================================================
Total params: 50,751
Trainable params: 50,751
Non-trainable params: 0
_________________________________________________________________

所以你的第一层有 2 个 dims (None,100)。错误消息说 LSTM 层需要 3 个维度，所以扩展 1 个暗淡：

model = keras.Sequential([
    keras.layers.Dense(100, activation='elu', input_shape=(481,)),
    keras.layers.Reshape((100,1)),
    keras.layers.LSTM(64, return_sequences=False, dropout=0.1, recurrent_dropout=0.1),
    keras.layers.Dense(25, activation='elu'),
    keras.layers.Dropout(0.5),
    keras.layers.Dense(1,activation= 'elu')
])

opt = keras.optimizers.Adam(lr=0.001)
model.compile(optimizer=opt, loss='mean_squared_error', metrics=['mse'])
model.summary()

你会得到：

Model: "sequential_5"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_16 (Dense)             (None, 100)               48200     
_________________________________________________________________
reshape_2 (Reshape)          (None, 100, 1)            0         
_________________________________________________________________
lstm_4 (LSTM)                (None, 64)                16896     
_________________________________________________________________
dense_17 (Dense)             (None, 25)                1625      
_________________________________________________________________
dropout_6 (Dropout)          (None, 25)                0         
_________________________________________________________________
dense_18 (Dense)             (None, 1)                 26        
=================================================================
Total params: 66,747
Trainable params: 66,747
Non-trainable params: 0
_________________________________________________________________

希望对你有所帮助。

如何为 LSTM 层的循环神经网络重塑张量

How to reshape tensor for Recurrent Neural Network for LSTM layer

python

keras

tensorflow

recurrent-neural-network