将LSTM与CNN连接，在Keras中具有不同的张量尺寸

Question

这是我试图通过使用连接操作合并的两个神经元网络。该网络应该通过1好和0坏电影对IMDB电影评论进行分类

def cnn_lstm_merged():
       embedding_vecor_length = 32
       cnn_model = Sequential()
       cnn_model.add(Embedding(top_words, embedding_vecor_length, input_length=max_review_length))
       cnn_model.add(Conv1D(filters=32, kernel_size=3, padding='same', activation='relu'))
       cnn_model.add(MaxPooling1D(pool_size=2))
       cnn_model.add(Flatten())

       lstm_model = Sequential()
       lstm_model.add(Embedding(top_words, embedding_vecor_length, input_length=max_review_length))
       lstm_model.add(LSTM(64, activation = 'relu'))
       lstm_model.add(Flatten())

       merge = concatenate([lstm_model, cnn_model])
       hidden = (Dense(1, activation = 'sigmoid'))(merge)
       #print(model.summary())
       output = hidden.fit(X_train, y_train, epochs=3, batch_size=64)
       return output

但是当我运行代码时出现错误：

  File "/home/pythonist/Desktop/EnsemblingLSTM_CONV/train.py", line 59, in cnn_lstm_merged
    lstm_model.add(Flatten())
  File "/home/pythonist/deeplearningenv/lib/python3.6/site-packages/keras/engine/sequential.py", line 185, in add
    output_tensor = layer(self.outputs[0])
  File "/home/pythonist/deeplearningenv/lib/python3.6/site-packages/keras/engine/base_layer.py", line 414, in __call__
    self.assert_input_compatibility(inputs)
  File "/home/pythonist/deeplearningenv/lib/python3.6/site-packages/keras/engine/base_layer.py", line 327, in assert_input_compatibility
    str(K.ndim(x)))
ValueError: Input 0 is incompatible with layer flatten_2: expected min_ndim=3, found ndim=2
[Finished in 4.8s with exit code 1]

如何合并这两层？谢谢

Answer 1

在Flatten之后没有必要使用LSTM，因为LSTM（默认情况下）只返回最后一个状态而不是序列，即数据将具有(BS, n_output)形状，但Flatten层需要(BS, a, b)形状将转换为(BS, a*b)。

因此，要么删除Flatten图层，只使用最后一个状态，要么将return_sequences=True添加到LSTM。这将使LSTM返回所有输出而不仅仅是最后一个输出，即(BS, T, n_out)。

编辑：此外，您创建最终模型的方式是错误的。请看看this的例子;对你来说，它应该是这样的：

  merge = Concatenate([lstm_model, cnn_model])
  hidden = Dense(1, activation = 'sigmoid')
  conc_model = Sequential()
  conc_model.add(merge)
  conc_model.add(hidden)
  conc_model.compile(...)

  output = conc_model .fit(X_train, y_train, epochs=3, batch_size=64)

总而言之，使用Functional API可能更好。

编辑2：这是最终的代码

cnn_model = Sequential()
cnn_model.add(Embedding(top_words, embedding_vecor_length, input_length=max_review_length))
cnn_model.add(Conv1D(filters=32, kernel_size=3, padding='same', activation='relu'))
cnn_model.add(MaxPooling1D(pool_size=2))
cnn_model.add(Flatten())

lstm_model = Sequential()
lstm_model.add(Embedding(top_words, embedding_vecor_length, input_length=max_review_length))
lstm_model.add(LSTM(64, activation = 'relu', return_sequences=True))
lstm_model.add(Flatten())

# instead of the last two lines you can also use
# lstm_model.add(LSTM(64, activation = 'relu'))
# then you do not have to use the Flatten layer. depends on your actual needs

merge = Concatenate([lstm_model, cnn_model])
hidden = Dense(1, activation = 'sigmoid')
conc_model = Sequential()
conc_model.add(merge)
conc_model.add(hidden)