def f(x):
return custom_gradient.prevent_gradient(x, 'No gradient')
))
]
return assertions
def _outer_squared_difference(x, y):
"""Convenience function analogous to tf.squared_difference."""
z = x - y
return z[..., tf.newaxis, :] * z[..., tf.newaxis]
@tfp_custom_gradient.custom_gradient(
vjp_fwd=lambda x: (tf.identity(x), ()),
vjp_bwd=lambda _, dx: tfp_custom_gradient.prevent_gradient( # pylint: disable=g-long-lambda
dx,
message='Second derivative is not implemented.'),
jvp_fn=lambda primals, tangents: ( # pylint: disable=g-long-lambda
tfp_custom_gradient.prevent_gradient(
primals[0], message='Second derivative is not implemented.'),
tfp_custom_gradient.prevent_gradient(
tangents[0], message='Second derivative is not implemented.')))
def _prevent_2nd_derivative(x):
"""Disables computation of the second derivatives for a tensor.
NB: you need to apply a non-identity function to the output tensor for the
exception to be raised.
Arguments:
x: A tensor.
def grad(dy):
return tfp_custom_gradient.prevent_gradient(
dy, message='Second derivative is not implemented.')