我试图使用Keras对单变量时间序列进行预测。

神经网络模型看起来像这样

Model: "sequential"_________________________________________________________________Layer (type)                 Output Shape              Param #   =================================================================conv1d (Conv1D)              (None, None, 25)          150       _________________________________________________________________lstm (LSTM)                  (None, None, 1024)        4300800   _________________________________________________________________dropout (Dropout)            (None, None, 1024)        0         _________________________________________________________________lstm_1 (LSTM)                (None, None, 1024)        8392704   _________________________________________________________________dropout_1 (Dropout)          (None, None, 1024)        0         _________________________________________________________________lstm_2 (LSTM)                (None, None, 1024)        8392704   _________________________________________________________________dropout_2 (Dropout)          (None, None, 1024)        0         _________________________________________________________________lstm_3 (LSTM)                (None, None, 1024)        8392704   _________________________________________________________________dropout_3 (Dropout)          (None, None, 1024)        0         _________________________________________________________________dense (Dense)                (None, None, 1)           1025      =================================================================Total params: 29,480,087Trainable params: 29,480,087Non-trainable params: 0_________________________________________________________________

我的数据使用系列的前三个值进行窗口化以预测下一个值。因此，我的测试数据集看起来像这样

list(dataset.as_numpy_iterator())[array([[[ 0.        ],         [ 0.02346429],         [ 0.04559132]],        [[ 0.        ],         [ 0.02161974],         [ 0.13014923]],        [[ 0.        ],         [ 0.10623277],         [-0.02918068]],        [[ 0.        ],         [-0.12240955],         [-0.21869095]]])]

一切正常，但是当我将这些数据输入到model.predict(dataset)时，输出的结果是

array([[[ 0.01316399],        [ 0.03728709],        [ 0.06164959]],       [[ 0.01316399],        [ 0.03512047],        [ 0.1292857 ]],       [[ 0.01316399],        [ 0.1172413 ],        [-0.01671433]],       [[ 0.01316399],        [-0.10654409],        [-0.16395506]]], dtype=float32)

这个例子中的形状是(4, 3, 1)

鉴于我的神经网络的最后一层是一个单元的Dense层，我期望每个输入特征三元组只得到一个预测结果。为什么每个训练样本的预测似乎有三个输出呢？

回答：

在你的最后一个LSTM层中，设置参数return_sequences=False。

LSTM(..., return_sequences=False)