def _tf_sparsemax_loss(self, z, q, dtype, use_gpu):
z = z.astype(dtype)
q = q.astype(dtype)
with self.test_session(use_gpu=use_gpu):
tf_sparsemax_op = sparsemax(z)
tf_loss_op = sparsemax_loss(z, tf_sparsemax_op, q)
tf_loss_out = tf_loss_op.eval()
return tf_loss_op, tf_loss_out
def _test_gradient_against_numpy(self, dtype, random, use_gpu):
"""check sparsemax Rop, aginst numpy Rop"""
z = random.uniform(low=-3, high=3, size=(test_obs, 10)).astype(dtype)
logits = constant_op.constant(z, name='z')
sparsemax_op = sparsemax(logits)
sparsemax_grad_op = gradients_impl.gradients(sparsemax_op, [logits])[0]
with self.test_session(use_gpu=use_gpu):
tf_grad = sparsemax_grad_op.eval()
np_grad = self._np_sparsemax_grad(z)
self.assertAllCloseAccordingToType(np_grad, tf_grad)
self.assertShapeEqual(np_grad, sparsemax_grad_op)
def _test_gradient_against_estimate(self, dtype, random, use_gpu):
"""check sparsemax Rop, aginst estimated Rop"""
z = random.uniform(low=-3, high=3, size=(test_obs, 10)).astype(dtype)
logits = array_ops.placeholder(dtype, name='z')
sparsemax_op = sparsemax(logits)
with self.test_session(use_gpu=use_gpu):
err = gradient_checker.compute_gradient_error(logits,
z.shape,
sparsemax_op,
z.shape,
x_init_value=z,
delta=1e-9)
self.assertLess(err, 1e-4)
def _test_gradient_against_numpy(self, dtype, random, use_gpu):
"""check sparsemax-loss Rop, aginst numpy Rop"""
z = random.uniform(low=-3, high=3, size=(test_obs, 10))
q = np.zeros((test_obs, 10))
q[np.arange(0, test_obs), np.random.randint(0, 10, size=test_obs)] = 1
logits = constant_op.constant(z.astype(dtype), name='z')
sparsemax_op = sparsemax(logits)
loss_op = sparsemax_loss(logits, sparsemax_op, q.astype(dtype))
loss_grad_op = gradients_impl.gradients(loss_op, [logits])[0]
with self.test_session(use_gpu=use_gpu):
tf_grad = loss_grad_op.eval()
np_grad = self._np_sparsemax_loss_grad(z, q).astype(dtype)
self.assertShapeEqual(np_grad, loss_grad_op)
def _test_gradient_against_estimate(self, dtype, random, use_gpu):
"""check sparsemax-loss Rop, aginst estimated-loss Rop"""
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
logits = array_ops.placeholder(dtype, name='z')
sparsemax_op = sparsemax(logits)
loss_op = sparsemax_loss(logits, sparsemax_op, q)
with self.test_session(use_gpu=use_gpu):
err = gradient_checker.compute_gradient_error(logits,
z.shape,
loss_op,
(test_obs, ),
x_init_value=z,
delta=1e-9)
self.assertLess(err, 1e-4)
def _tf_sparsemax(self, z, dtype, use_gpu):
with self.test_session(use_gpu=use_gpu):
tf_sparsemax_op = sparsemax(z.astype(dtype))
tf_sparsemax_out = tf_sparsemax_op.eval()
return tf_sparsemax_op, tf_sparsemax_out