def test_get_score(self):
class Capture(object):
def __init__(self, vae):
self._vae = vae
self._vae_variational = vae.variational
vae.variational = self._variational
self.q_net = None
def _variational(self, x, z=None, n_z=None):
self.q_net = self._vae_variational(x, z=z, n_z=n_z)
return self.q_net
tf.set_random_seed(1234)
donut = Donut(
h_for_p_x=lambda x: x,
h_for_q_z=lambda x: x,
x_dims=5,
z_dims=3,
)
capture = Capture(donut.vae)
donut.vae.reconstruct = Mock(
wraps=lambda x: x + tf.reduce_sum(x) # only called by MCMC
)
x = tf.reshape(tf.range(20, dtype=tf.float32), [4, 5])
y = tf.constant([[1, 0, 0, 0, 0], [0, 1, 0, 0, 0], [0, 0, 1, 0, 0],
[0, 0, 0, 1, 0]],
dtype=tf.int32)
_ = donut.get_score(x, y) # ensure variables created
def r_prob(x, z, n_z=None, x_in=None):
if x_in is None:
x_in = x
q_net = donut.vae.variational(x_in, z=z, n_z=n_z)
p_net = donut.vae.model(z=q_net['z'], x=x, n_z=n_z)
p = p_net['x'].log_prob(group_ndims=0)
if n_z is not None:
p = tf.reduce_mean(p, axis=0)
return p
with self.test_session() as sess:
ensure_variables_initialized()
# test y is None
donut.vae.reconstruct.reset_mock()
np.testing.assert_allclose(*sess.run([
donut.get_score(
x, y=None, mcmc_iteration=1, last_point_only=False),
r_prob(x, z=capture.q_net['z'])
]))
self.assertEqual(donut.vae.reconstruct.call_count, 0)
# test mcmc_iteration is None
donut.vae.reconstruct.reset_mock()
np.testing.assert_allclose(*sess.run([
donut.get_score(
x, y=y, mcmc_iteration=None, last_point_only=False),
r_prob(x, z=capture.q_net['z'])
]))
self.assertEqual(donut.vae.reconstruct.call_count, 0)
# test mcmc once
x2 = tf.where(
tf.cast(y, dtype=tf.bool),
x,
x + tf.reduce_sum(x),
)
donut.vae.reconstruct.reset_mock()
np.testing.assert_allclose(*sess.run([
donut.get_score(
x, y=y, mcmc_iteration=1, last_point_only=False),
r_prob(x, z=capture.q_net['z'], x_in=x2)
]))
self.assertEqual(donut.vae.reconstruct.call_count, 1)
# test mcmc with n_z > 1
donut.vae.reconstruct.reset_mock()
np.testing.assert_allclose(*sess.run([
donut.get_score(
x, y=y, n_z=7, mcmc_iteration=1, last_point_only=False),
r_prob(x, z=capture.q_net['z'], x_in=x2, n_z=7)
]))
self.assertEqual(capture.q_net['z'].get_shape(),
tf.TensorShape([7, 4, 3]))
self.assertEqual(donut.vae.reconstruct.call_count, 1)
def test_prediction(self):
np.random.seed(1234)
tf.set_random_seed(1234)
# test last_point_only == True
donut = Donut(h_for_p_x=lambda x: x,
h_for_q_z=lambda x: x,
x_dims=5,
z_dims=3)
_ = donut.get_score(tf.zeros([4, 5], dtype=tf.float32),
tf.zeros(
[4, 5],
dtype=tf.int32)) # ensure variables created
pred = DonutPredictor(donut, n_z=2, batch_size=4)
self.assertIs(pred.model, donut)
with self.test_session():
ensure_variables_initialized()
# test without missing
res = pred.get_score(values=np.arange(5, dtype=np.float32))
self.assertEqual(res.shape, (1, ))
res = pred.get_score(values=np.arange(8, dtype=np.float32))
self.assertEqual(res.shape, (4, ))
res = pred.get_score(values=np.arange(10, dtype=np.float32))
self.assertEqual(res.shape, (6, ))
# test with missing
res = pred.get_score(values=np.arange(10, dtype=np.float32),
missing=np.random.binomial(1, .5, size=10))
self.assertEqual(res.shape, (6, ))
# test errors
with self.test_session():
with pytest.raises(ValueError,
match='`values` must be a 1-D array'):
_ = pred.get_score(
np.arange(10, dtype=np.float32).reshape([-1, 1]))
with pytest.raises(ValueError,
match='The shape of `missing` does not agree '
'with the shape of `values`'):
_ = pred.get_score(np.arange(10, dtype=np.float32),
np.arange(9, dtype=np.int32))
# test last_point_only == False
pred = DonutPredictor(donut,
n_z=2,
batch_size=4,
last_point_only=False)
with self.test_session():
ensure_variables_initialized()
# test without missing
res = pred.get_score(values=np.arange(10, dtype=np.float32))
self.assertEqual(res.shape, (6, 5))
# test with missing
res = pred.get_score(values=np.arange(10, dtype=np.float32),
missing=np.random.binomial(1, .5, size=10))
self.assertEqual(res.shape, (6, 5))
# test set feed_dict
is_training = tf.placeholder(tf.bool, shape=())
pred = DonutPredictor(donut,
n_z=2,
batch_size=4,
feed_dict={is_training: False})
with self.test_session():
ensure_variables_initialized()
_ = pred.get_score(values=np.arange(10, dtype=np.float32))