创建级联神经网络以及如何训练它

Question

我有一个关于级联神经网络的输入形状的问题。我有一个标记化的文本列，使用函数 pad_sequences 所有数据点的长度为 2395，用于训练我有 6493 个数据点。所以文本部分的形状是 (6493, 2395)，不是吗？我有 17 个附加列要放入模型中。所以这个附加数据的形状是 (6493, 17)。

对于神经网络，我有以下代码：

embedding_dim = 300
inp_dim = X_train.shape[1]
text_data = Input(shape=(max_length,), name="X_train")
meta_data = Input(shape=X_train_zusatz.shape, name="X_train_zusatz")

x1 = (Embedding(input_dim=vocab_size, output_dim=embedding_dim, input_length=max_length))(text_data)
x2 = (LSTM(300, dropout = 0.2, recurrent_dropout = 0.2, return_sequences=True))(x1)
x3 = (Dense(300, activation = "relu"))(meta_data)
x4 = concatenate([x2, x3], axis = 1)
x5 = (Dense(300, activation = "relu"))(x4)
x6 = Dropout(0.25)(x5)
x7 = (Dense(300, activation = "relu"))(x6)
x8 = BatchNormalization()(x7)
x9 = (Dense(4, activation='softmax'))(x8)

model = Model(inputs = [text_data, meta_data], outputs = x9)
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

print(model.summary())

model.summary 如下所示：

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
X_train (InputLayer)            (None, 2395)         0                                            
__________________________________________________________________________________________________
embedding_19 (Embedding)        (None, 2395, 300)    23400600    X_train[0][0]                    
__________________________________________________________________________________________________
X_train_zusatz (InputLayer)     (None, 6493, 17)     0                                            
__________________________________________________________________________________________________
lstm_19 (LSTM)                  (None, 2395, 300)    721200      embedding_19[0][0]               
__________________________________________________________________________________________________
dense_44 (Dense)                (None, 6493, 300)    5400        X_train_zusatz[0][0]             
__________________________________________________________________________________________________
concatenate_18 (Concatenate)    (None, 8888, 300)    0           lstm_19[0][0]                    
                                                                 dense_44[0][0]                   
__________________________________________________________________________________________________
dense_45 (Dense)                (None, 8888, 300)    90300       concatenate_18[0][0]             
__________________________________________________________________________________________________
dropout_1 (Dropout)             (None, 8888, 300)    0           dense_45[0][0]                   
__________________________________________________________________________________________________
dense_46 (Dense)                (None, 8888, 300)    90300       dropout_1[0][0]                  
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 8888, 300)    1200        dense_46[0][0]                   
__________________________________________________________________________________________________
dense_47 (Dense)                (None, 8888, 4)      1204        batch_normalization_7[0][0]      
==================================================================================================
Total params: 24,310,204
Trainable params: 24,309,604
Non-trainable params: 600
__________________________________________________________________________________________________
None

所以我的问题是，为什么我在嵌入层之后看不到数据点的数量（6493）。我在这一层犯了什么错误吗？因为在密集层中我得到了形状 (None, 6493, 300)，但在嵌入层中我得到了 (None, 2395, 300)。恐怕这里的列和行是混在一起的吧？

除此之外，我无法训练模型。代码：

model.fit([X_train, X_train_zusatz], y_train, epochs=100, batch_size=500, validation_data=[[X_test, X_test_zusatz], y_test], class_weight=class_weight)

会导致错误：

ValueError: Error when checking input: expected X_train_zusatz to have 3 dimensions, but got array with shape (6493, 17)

我该如何解决？因为 (6493, 17) 是附加数据的正确形状，但我的神经网络不会接受它。

非常感谢！

最好的问候，丹尼尔

Answer 1

您正在使用需要 3 维输入的 LSTM 层，但您知道您的输入是 2 维的。不要使用 LSTM，而是使用 Dense。