Keras 密集输入大小错误，展平返回（无，无）

Question

I'm trying to implement this model to generate midi music but I'm getting an error

The last dimension of the inputs to `Dense` should be defined. Found `None`.

Here's my code

model = Sequential()
model.add(Bidirectional(LSTM(512, return_sequences=True), input_shape=(network_input.shape[1], network_input.shape[2])))
model.add(SeqSelfAttention(attention_activation='sigmoid'))
model.add(Dropout(0.3))
model.add( LSTM(512, return_sequences=True))
model.add(Dropout(0.3))
model.add(Flatten())
model.summary()
model.add(Dense(note_variants_count))
model.compile(loss='categorical_crossentropy', optimizer='adam')

And here's the summary before the dense layer

Model: "sequential_17"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
bidirectional_19 (Bidirectio (None, 100, 1024)         2105344   
_________________________________________________________________
seq_self_attention_20 (SeqSe (None, None, 1024)        65601     
_________________________________________________________________
dropout_38 (Dropout)         (None, None, 1024)        0         
_________________________________________________________________
lstm_40 (LSTM)               (None, None, 512)         3147776   
_________________________________________________________________
dropout_39 (Dropout)         (None, None, 512)         0         
_________________________________________________________________
flatten_14 (Flatten)         (None, None)              0         
=================================================================
Total params: 5,318,721
Trainable params: 5,318,721
Non-trainable params: 0
_________________________________________________________________

I think that the Flatten layer is causing the problem but I have no idea why it's returning a (None, None) shape.

Answer 1

您必须为 Flatten 层提供除批次维度之外的所有维度。

Answer 2

Using RNN and 'return_sequence'

For RNNs like LSTM, there is an option to either return the whole sequence or just the results. In this case you want just the results. Changing just this one line

model.add(Dropout(0.3))
model.add(LSTM(512, return_sequences=False)) # <== change to False
model.add(Dropout(0.3))
model.add(Flatten())
model.summary()

Returns the following network shape from summary

_________________________________________________________________
lstm_4 (LSTM)                (None, 512)               3147776   
_________________________________________________________________
dropout_3 (Dropout)          (None, 512)               0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 512)               0         
=================================================================

Notice the reduction of rank of the output tensor from Rank 3 to Rank 2. This is because this output is just that, the output, and not the whole sequence under consideration with all the hidden states.