def test_kl_with_logits(self):
p_logits = tf.placeholder(tf.float32, shape=(100, 20))
q_logits = tf.placeholder(tf.float32, shape=(100, 20))
p_logits_np = np.random.normal(0, 10, size=(100, 20))
q_logits_np = np.random.normal(0, 10, size=(100, 20))
kl_div_tf = self.sess.run(utils_tf.kl_with_logits(p_logits, q_logits),
feed_dict={p_logits: p_logits_np,
q_logits: q_logits_np})
kl_div_ref = numpy_kl_with_logits(p_logits_np, q_logits_np)
self.assertClose(kl_div_ref, kl_div_tf)
def test_kl_with_logits(self):
p_logits = tf.placeholder(tf.float32, shape=(100, 20))
q_logits = tf.placeholder(tf.float32, shape=(100, 20))
p_logits_np = np.random.normal(0, 10, size=(100, 20))
q_logits_np = np.random.normal(0, 10, size=(100, 20))
with tf.Session() as sess:
kl_div_tf = sess.run(utils_tf.kl_with_logits(p_logits, q_logits),
feed_dict={p_logits: p_logits_np,
q_logits: q_logits_np})
kl_div_ref = numpy_kl_with_logits(p_logits_np, q_logits_np)
self.assertTrue(np.allclose(kl_div_ref, kl_div_tf))
def test_kl_with_logits(self):
import tensorflow as tf
q_logits = tf.placeholder(tf.float32, shape=(100, 20))
p_logits = tf.placeholder(tf.float32, shape=(100, 20))
q_logits_np = np.random.normal(0, 10, size=(100, 20))
p_logits_np = np.random.normal(0, 10, size=(100, 20))
with tf.Session() as sess:
kl_div_tf = sess.run(kl_with_logits(q_logits, p_logits),
feed_dict={
q_logits: q_logits_np,
p_logits: p_logits_np
})
kl_div_ref = numpy_kl_with_logits(q_logits_np, p_logits_np)
self.assertTrue(np.allclose(kl_div_ref, kl_div_tf))
def vatm(
model,
x,
logits,
eps,
num_iterations=1,
xi=1e-6,
clip_min=None,
clip_max=None,
scope=None,
):
"""
Tensorflow implementation of the perturbation method used for virtual
adversarial training: https://arxiv.org/abs/1507.00677
:param model: the model which returns the network unnormalized logits
:param x: the input placeholder
:param logits: the model's unnormalized output tensor (the input to
the softmax layer)
:param eps: the epsilon (input variation parameter)
:param num_iterations: the number of iterations
:param xi: the finite difference parameter
:param clip_min: optional parameter that can be used to set a minimum
value for components of the example returned
:param clip_max: optional parameter that can be used to set a maximum
value for components of the example returned
:param seed: the seed for random generator
:return: a tensor for the adversarial example
"""
with tf.name_scope(scope, "virtual_adversarial_perturbation"):
d = tf.random_normal(tf.shape(x), dtype=tf_dtype)
for _ in range(num_iterations):
d = xi * utils_tf.l2_batch_normalize(d)
logits_d = model.get_logits(x + d)
kl = utils_tf.kl_with_logits(logits, logits_d)
Hd = tf.gradients(kl, d)[0]
d = tf.stop_gradient(Hd)
d = eps * utils_tf.l2_batch_normalize(d)
adv_x = x + d
if (clip_min is not None) and (clip_max is not None):
adv_x = tf.clip_by_value(adv_x, clip_min, clip_max)
return adv_x
