def _test(distribution, bijector, n):
x = TransformedDistribution(distribution=distribution,
bijector=bijector,
validate_args=True)
val_est = get_dims(x.sample(n))
val_true = n + get_dims(distribution.mean())
assert val_est == val_true
def test_auto_transform_true(self):
with self.test_session() as sess:
# Match normal || softplus-inverse-normal distribution with
# automated transformation on latter (assuming it is softplus).
x = TransformedDistribution(
distribution=Normal(0.0, 0.5),
bijector=tf.contrib.distributions.bijectors.Softplus())
x.support = 'nonnegative'
qx = Normal(loc=tf.Variable(tf.random_normal([])),
scale=tf.nn.softplus(tf.Variable(tf.random_normal(
[]))))
inference = ed.KLqp({x: qx})
inference.initialize(auto_transform=True, n_samples=5, n_iter=1000)
tf.global_variables_initializer().run()
for _ in range(inference.n_iter):
info_dict = inference.update()
# Check approximation on constrained space has same moments as
# target distribution.
n_samples = 10000
x_mean, x_var = tf.nn.moments(x.sample(n_samples), 0)
x_unconstrained = inference.transformations[x]
qx_constrained = transform(
qx, bijectors.Invert(x_unconstrained.bijector))
qx_mean, qx_var = tf.nn.moments(qx_constrained.sample(n_samples),
0)
stats = sess.run([x_mean, qx_mean, x_var, qx_var])
self.assertAllClose(info_dict['loss'], 0.0, rtol=0.2, atol=0.2)
self.assertAllClose(stats[0], stats[1], rtol=1e-1, atol=1e-1)
self.assertAllClose(stats[2], stats[3], rtol=1e-1, atol=1e-1)
def test_auto_transform_true(self):
with self.test_session() as sess:
# Match normal || softplus-inverse-normal distribution with
# automated transformation on latter (assuming it is softplus).
x = TransformedDistribution(
distribution=Normal(0.0, 0.5),
bijector=tf.contrib.distributions.bijectors.Softplus())
x.support = 'nonnegative'
qx = Normal(loc=tf.Variable(tf.random_normal([])),
scale=tf.nn.softplus(tf.Variable(tf.random_normal([]))))
inference = ed.KLqp({x: qx})
inference.initialize(auto_transform=True, n_samples=5, n_iter=1000)
tf.global_variables_initializer().run()
for _ in range(inference.n_iter):
info_dict = inference.update()
# Check approximation on constrained space has same moments as
# target distribution.
n_samples = 10000
x_mean, x_var = tf.nn.moments(x.sample(n_samples), 0)
x_unconstrained = inference.transformations[x]
qx_constrained = transform(qx, bijectors.Invert(x_unconstrained.bijector))
qx_mean, qx_var = tf.nn.moments(qx_constrained.sample(n_samples), 0)
stats = sess.run([x_mean, qx_mean, x_var, qx_var])
self.assertAllClose(info_dict['loss'], 0.0, rtol=0.2, atol=0.2)
self.assertAllClose(stats[0], stats[1], rtol=1e-1, atol=1e-1)
self.assertAllClose(stats[2], stats[3], rtol=1e-1, atol=1e-1)
def test_hmc_default(self):
with self.test_session() as sess:
x = TransformedDistribution(
distribution=Normal(1.0, 1.0),
bijector=tf.contrib.distributions.bijectors.Softplus())
x.support = 'nonnegative'
inference = ed.HMC([x])
inference.initialize(auto_transform=True, step_size=0.8)
tf.global_variables_initializer().run()
for _ in range(inference.n_iter):
info_dict = inference.update()
inference.print_progress(info_dict)
# Check approximation on constrained space has same moments as
# target distribution.
n_samples = 10000
x_unconstrained = inference.transformations[x]
qx = inference.latent_vars[x_unconstrained]
qx_constrained = Empirical(
x_unconstrained.bijector.inverse(qx.params))
x_mean, x_var = tf.nn.moments(x.sample(n_samples), 0)
qx_mean, qx_var = tf.nn.moments(qx_constrained.params[500:], 0)
stats = sess.run([x_mean, qx_mean, x_var, qx_var])
self.assertAllClose(stats[0], stats[1], rtol=1e-1, atol=1e-1)
self.assertAllClose(stats[2], stats[3], rtol=1e-1, atol=1e-1)
def _test(base_dist_cls, transform, inverse, log_det_jacobian, n,
**base_dist_args):
x = TransformedDistribution(base_dist_cls=base_dist_cls,
transform=transform,
inverse=inverse,
log_det_jacobian=log_det_jacobian,
**base_dist_args)
val_est = get_dims(x.sample(n))
val_true = n + get_dims(base_dist_args['mu'])
assert val_est == val_true
def _test(base_dist_cls, transform, inverse, log_det_jacobian, n, **base_dist_args):
x = TransformedDistribution(
base_dist_cls=base_dist_cls,
transform=transform,
inverse=inverse,
log_det_jacobian=log_det_jacobian,
**base_dist_args
)
val_est = get_dims(x.sample(n))
val_true = n + get_dims(base_dist_args["mu"])
assert val_est == val_true
def test_hmc_default(self):
with self.test_session() as sess:
x = TransformedDistribution(
distribution=Normal(1.0, 1.0),
bijector=tf.contrib.distributions.bijectors.Softplus())
x.support = 'nonnegative'
inference = ed.HMC([x])
inference.initialize(auto_transform=True, step_size=0.8)
tf.global_variables_initializer().run()
for _ in range(inference.n_iter):
info_dict = inference.update()
inference.print_progress(info_dict)
# Check approximation on constrained space has same moments as
# target distribution.
n_samples = 1000
qx_constrained = inference.latent_vars[x]
x_mean, x_var = tf.nn.moments(x.sample(n_samples), 0)
qx_mean, qx_var = tf.nn.moments(qx_constrained.params[500:], 0)
stats = sess.run([x_mean, qx_mean, x_var, qx_var])
self.assertAllClose(stats[0], stats[1], rtol=1e-1, atol=1e-1)
self.assertAllClose(stats[2], stats[3], rtol=1e-1, atol=1e-1)
def test_hmc_custom(self):
with self.test_session() as sess:
x = TransformedDistribution(
distribution=Normal(1.0, 1.0),
bijector=tf.contrib.distributions.bijectors.Softplus())
x.support = 'nonnegative'
qx = Empirical(tf.Variable(tf.random_normal([1000])))
inference = ed.HMC({x: qx})
inference.initialize(auto_transform=True, step_size=0.8)
tf.global_variables_initializer().run()
for _ in range(inference.n_iter):
info_dict = inference.update()
# Check approximation on constrained space has same moments as
# target distribution.
n_samples = 10000
x_unconstrained = inference.transformations[x]
qx_constrained_params = x_unconstrained.bijector.inverse(qx.params)
x_mean, x_var = tf.nn.moments(x.sample(n_samples), 0)
qx_mean, qx_var = tf.nn.moments(qx_constrained_params[500:], 0)
stats = sess.run([x_mean, qx_mean, x_var, qx_var])
self.assertAllClose(stats[0], stats[1], rtol=1e-1, atol=1e-1)
self.assertAllClose(stats[2], stats[3], rtol=1e-1, atol=1e-1)
n_rep = 100 # number of replicated datasets we generate
holdout_gen = np.zeros((n_rep, x_train.shape[0], x_train.shape[1]))
for i in range(n_rep):
x_generated = x_post.sample().eval()
# look only at the heldout entries
holdout_gen[i] = np.multiply(x_generated, holdout_mask)
n_eval = 10 # we draw samples from the inferred Z and W
obs_ll = []
rep_ll = []
pbar = Progbar(n_eval)
for j in range(n_eval):
U_sample = U_post.sample().eval()
V_sample = V_post.sample().eval()
holdoutmean_sample = np.multiply(U_sample.dot(V_sample.T), holdout_mask)
obs_ll.append(
np.mean(np.ma.masked_invalid(
stats.poisson.logpmf(np.array(x_vad, dtype=int),
holdoutmean_sample)),
axis=1))
rep_ll.append(
np.mean(np.ma.masked_invalid(
stats.poisson.logpmf(holdout_gen, holdoutmean_sample)),
axis=2))
pbar.update(j)
