tf. keras. layers. Timedistributed function

The function prototype

tf.keras.layers.TimeDistributed(layer, 
								**kwargs
)

Function description

The time distribution layer is mainly used to slice the time dimension of the input data . At each time step , Enter one item in turn , And output one item in turn .
In the diagram above , The function of the time distribution layer is in time t input data w, Output data x; In time t1 input data x, Output data y. Enter... In turn in each time slice 、 Output a data , Instead of entering multiple items directly , Output multiple .

Parameters layer For one tf.keras.layers.Layer example . The input data is at least one three-dimensional tensor , If the input data is text sequence data , The input is a three-dimensional tensor , Shape is (batch_size, time, output_dim), If the input data is video data , The input is a five-dimensional tensor , Shape is (batch_size, time, height, width, channels).

The time distribution layer is used to process data with time dimension , It is not applicable to data without time dimension such as images .

Function USES

model = tf.keras.models.Sequential([
    #  Embedded layer 
    tf.keras.layers.Embedding(len(words), 125),
    # LSTM layer 
    tf.keras.layers.LSTM(125, return_sequences=True),
    # LSTM layer 
    tf.keras.layers.LSTM(125, return_sequences=True),
    #  Fully connected in the time dimension Dense layer 
    tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(len(words), activation="softmax"))
])

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 embedding (Embedding)       (None, None, 125)         32000     
                                                                 
 lstm (LSTM)                 (None, None, 125)         125500    
                                                                 
 lstm_1 (LSTM)               (None, None, 125)         125500    
                                                                 
 time_distributed (TimeDistr  (None, None, 256)        32256     
 ibuted)                                                         
                                                                 
=================================================================
Total params: 315,256
Trainable params: 315,256
Non-trainable params: 0
_________________________________________________________________