tf. keras. layers. Inputlayer function

The function prototype

tf.keras.layers.InputLayer(input_shape=None, 
						   batch_size=None, 
						   dtype=None, 
						   input_tensor=None, 
						   sparse=None, 
						   name=None, 
						   ragged=None, 
						   type_spec=None, 
						   **kwargs
)

Function USES

The input layer is generally used as the first layer of the linear model , By specifying parameters input_tensor Wrap existing tensors or specify parameters input_shape Create a new placeholder tensor .

tf.compat.v1.disable_eager_execution()
a = tf.compat.v1.placeholder("float", (None, 32))
model = tf.keras.Sequential([
    tf.keras.layers.InputLayer(input_tensor=a),
    tf.keras.layers.Dense(64)
])

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense (Dense)               (None, 64)                2112      
                                                                 
=================================================================
Total params: 2,112
Trainable params: 2,112
Non-trainable params: 0
_________________________________________________________________

First, a placeholder tensor is created a, Shape is (None, 32). And then by specifying input_tensor=a Package tensor a As input tensor . The final output tensor shape is (None, 64).

model = tf.keras.Sequential([
    tf.keras.layers.InputLayer(input_shape=32),
    tf.keras.layers.Dense(64)
])

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense (Dense)               (None, 64)                2112      
                                                                 
=================================================================
Total params: 2,112
Trainable params: 2,112
Non-trainable params: 0
_________________________________________________________________

Specify parameters input_shape=32, Create a placeholder tensor tensor As input tensor , Shape is (None, 32). Through a Dense layer , The final output tensor shape is also (None, 64).

It is worth noting that , If you also specify input_shape and input_tensor, that input_shape Parameter will fail , That is to create an input layer based on the existing tensor .

You can also specify... On other layers input_shape Parameter to omit the layer .

model = tf.keras.Sequential([
    tf.keras.layers.Dense(64, input_shape=(None, 32))
])

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense (Dense)               (None, None, 64)          2112      
                                                                 
=================================================================
Total params: 2,112
Trainable params: 2,112
Non-trainable params: 0
_________________________________________________________________

By means of Dense Specify on layer input_shape=(None, 32), It's equivalent to creating a tensor with a shape of （None, None, 32） The input layer of . adopt Dense After the layer , The final shape is (None, None, 64).

tf.keras.layers.InputLayer Functions and tf.keras.Input function difference ：

tf.keras.Input The function returns a tensor Tensor,tf.keras.layers.InputLayer Function is a layer object . The suggestion on the official website is to use tf.keras.Input Function to create a InputLayer Input layer , Instead of using it directly tf.keras.layers.InputLayer To define the input layer .