def _tf_sparsemax_loss(self, z, q, dtype):
z = z.astype(dtype)
q = q.astype(dtype)
tf_sparsemax_op = sparsemax(z)
tf_loss_op = sparsemax_loss(z, tf_sparsemax_op, q)
tf_loss_out = self.evaluate(tf_loss_op)
return tf_loss_op, tf_loss_out
def _tf_sparsemax_loss(z, q, dtype):
z = z.astype(dtype)
q = q.astype(dtype)
tf_sparsemax_op = sparsemax(z)
tf_loss_op = sparsemax_loss(z, tf_sparsemax_op, q)
tf_loss_out = tf_loss_op
return tf_loss_op, tf_loss_out
def test_gradient_against_estimate(self, dtype=None):
"""check sparsemax-loss Rop, against estimated-loss Rop."""
random = np.random.RandomState(7)
# sparsemax is not a smooth function so gradient estimation is only
# possible for float64.
if dtype != "float64":
return
z = random.uniform(low=-3, high=3, size=(test_obs, 10)).astype(dtype)
q = np.zeros((test_obs, 10)).astype(dtype)
q[np.arange(0, test_obs), np.random.randint(0, 10, size=test_obs)] = 1
(jacob_sym, ), (jacob_num, ) = tf.test.compute_gradient(
lambda logits: sparsemax_loss(logits, sparsemax(logits), q), [z])
self.assertAllCloseAccordingToType(jacob_sym, jacob_num)
