Keras official Chinese document: Wrapper wrapper

Wrapper wrappertimedistributed Packaging Device

keras.layers.wrappers.TimeDistributed(layer)

The wrapper can apply a layer to each time step of the input

Parameters

• Layer:keras Layer Object

Entering a dimension of at least 3D and subscript 1 will be considered a time dimension

For example, consider a batch with 32 samples, each of which is a sequence of 10 vectors, each with a length of 16, the input dimension is (32,10,16) , it does not contain batch size input_shape for(10,16)

We can use the wrapper TimeDistributed wrapper Dense to produce a separate full connection for each time step signal:

# as the first layer in a modelmodel = Sequential()model.add(TimeDistributed(Dense(8), input_shape=(10, 16)))# now model.output_shape == (None, 10, 8)# subsequent layers: no need for input_shapemodel.add(TimeDistributed(Dense(32)))# now model.output_shape == (None, 10, 32)

The output data of the program is shape(32,10,8)

TimeDistributedthe use Dense of packaging is strictly equivalent to layers.TimeDistribuedDense . The difference is that the wrapper TimeDistribued can also be packaged on other layers, such as the Convolution2D packaging:

model = Sequential()model.add(TimeDistributed(Convolution2D(64, 3, 3), input_shape=(10, 3, 299, 299)))

Bidirectional packaging Device

keras.layers.wrappers.Bidirectional(layer, merge_mode=‘concat‘, weights=None)

Bidirectional RNN Wrapper

Parameters

• Layer: Recurrent Object
• Merge_mode: The combination of forward and back rnn outputs, for,, sum , mul concat ave and None one, if set to None, the return value is not combined, but is returned as a list

Example

model = Sequential()model.add(Bidirectional(LSTM(10, return_sequences=True), input_shape=(5, 10)))model.add(Bidirectional(LSTM(10)))model.add(Dense(5))model.add(Activation(‘softmax‘))model.compile(loss=‘categorical_crossentropy‘, optimizer=‘rmsprop‘)

Albert (http://www.aibbt.com/) The first artificial intelligence portal in China

Keras official Chinese document: Wrapper wrapper