def test_calc_kl_analytical(self):
hparams = hparams_mod.make_hparams(use_monte_carlo_kl=False)
dist_a = tf.distributions.Bernoulli(probs=0.5)
dist_b = tf.distributions.Bernoulli(probs=0.3)
kl_div = util.calc_kl(hparams, dist_a.sample(), dist_a, dist_b)
with self.test_session():
self.assertAllClose(kl_div.eval(),
0.5 * (np.log(0.5 / 0.3) + np.log(0.5 / 0.7)))
def _test_vae(self, vae_type):
"""Make sure that all tensors and assertions evaluate without error."""
hparams = hparams_mod.make_hparams(vae_type=vae_type)
inputs, vae = _inputs_and_vae(hparams)
tensors = _all_tensors(hparams, inputs, vae)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tensors)
def test_calc_kl_mc(self):
tf.set_random_seed(0)
hparams = hparams_mod.make_hparams(use_monte_carlo_kl=True)
samples = 1000
dist_a = tf.distributions.Bernoulli(probs=tf.fill([samples], 0.5))
dist_b = tf.distributions.Bernoulli(probs=tf.fill([samples], 0.3))
kl_div = tf.reduce_mean(util.calc_kl(hparams, dist_a.sample(), dist_a,
dist_b),
axis=0)
with self.test_session():
self.assertAllClose(kl_div.eval(),
0.5 * (np.log(0.5 / 0.3) + np.log(0.5 / 0.7)),
atol=0.05)
def _setup_model(self, session_params):
self.train_dataset = "train"
self.valid_dataset = "valid"
self.hparams = hparams_mod.make_hparams()
self.model = MockModel(self.hparams, session_params)
def make_hparams(flag_value=None, **kwargs):
"""Initialize HParams with the defaults in this module."""
init = dict(_DEFAULTS)
init.update(kwargs)
ret = hparams_mod.make_hparams(flag_value=flag_value, **init)
return ret