Gotcha: Tapes are single-use by default (Personal website)

Up: Automatic differentiation with GradientTape [Top][Contents]

Gotcha: Tapes are single-use by default

When GradientTape.gradient is called, all resources are released.

To compute multiple gradients, use multiple tapes or:

import tensorflow as tf

x1 = tf.Variable(1.)
x2 = tf.Variable(2.)

with tf.GradientTape(persistent=True) as tape:
    y1 = x1 ** 2
    y2 = x2 ** 2

dy_dx_1 = tape.gradient(y1, x1)
dy_dx_2 = tape.gradient(y2, x2)

del tape # NOTE Do not forget.

f"{dy_dx_1}, {dy_dx_2}"

2.0, 4.0