tf.identity

Return a Tensor with the same shape and contents as input.

tf.identity(
    input, name=None
)

Used in the notebooks

The return value is not the same Tensor as the original, but contains the same values. This operation is fast when used on the same device.

For example:

a = tf.constant([0.78])
a_identity = tf.identity(a)
a.numpy()
array([0.78], dtype=float32)
a_identity.numpy()
array([0.78], dtype=float32)

Calling tf.identity on a variable will make a Tensor that represents the value of that variable at the time it is called. This is equivalent to calling <variable>.read_value().

a = tf.Variable(5)
a_identity = tf.identity(a)
a.assign_add(1)
<tf.Variable ... shape=() dtype=int32, numpy=6>
a.numpy()
6
a_identity.numpy()
5

This function can also be used to explicitly transfer tensors between devices. For example, to transfer a tensor in GPU memory back to host memory, one can use:

with tf.device("/gpu:0"):
  x_on_gpu = tf.constant(1)
with tf.device("/cpu:0"):
  x_on_cpu = tf.identity(x_on_gpu)
x_on_cpu.device
'/job:localhost/replica:0/task:0/device:CPU:0'