def main(_):
dp = dirichlet_process(10.0)
# The number of sticks broken is dynamic, changing across evaluations.
sess = tf.Session()
print(sess.run(dp))
print(sess.run(dp))
# Demo of the DirichletProcess random variable in Edward.
base = Normal(0.0, 1.0)
# Highly concentrated DP.
alpha = 1.0
dp = DirichletProcess(alpha, base)
x = dp.sample(1000)
samples = sess.run(x)
plt.hist(samples, bins=100, range=(-3.0, 3.0))
plt.title("DP({0}, N(0, 1))".format(alpha))
plt.show()
# More spread out DP.
alpha = 50.0
dp = DirichletProcess(alpha, base)
x = dp.sample(1000)
samples = sess.run(x)
plt.hist(samples, bins=100, range=(-3.0, 3.0))
plt.title("DP({0}, N(0, 1))".format(alpha))
plt.show()
# States persist across calls to sample() in a DP.
alpha = 1.0
dp = DirichletProcess(alpha, base)
x = dp.sample(50)
y = dp.sample(75)
samples_x, samples_y = sess.run([x, y])
plt.subplot(211)
plt.hist(samples_x, bins=100, range=(-3.0, 3.0))
plt.title("DP({0}, N(0, 1)) across two calls to sample()".format(alpha))
plt.subplot(212)
plt.hist(samples_y, bins=100, range=(-3.0, 3.0))
plt.show()
# `theta` is the distribution indirectly returned by the DP.
# Fetching theta is the same as fetching the Dirichlet process.
dp = DirichletProcess(alpha, base)
theta = Normal(0.0, 1.0, value=tf.cast(dp, tf.float32))
print(sess.run([dp, theta]))
print(sess.run([dp, theta]))
# DirichletProcess can also take in non-scalar concentrations and bases.
alpha = tf.constant([0.1, 0.6, 0.4])
base = Exponential(rate=tf.ones([5, 2]))
dp = DirichletProcess(alpha, base)
print(dp)
def test_persistent_state(self):
with self.test_session() as sess:
dp = DirichletProcess(0.1, Normal(loc=0.0, scale=1.0))
x = dp.sample(5)
y = dp.sample(5)
x_data, y_data, locs = sess.run([x, y, dp.locs])
for sample in x_data:
self.assertTrue(sample in locs)
for sample in y_data:
self.assertTrue(sample in locs)
def test_persistent_state(self):
with self.test_session() as sess:
dp = DirichletProcess(0.1, Normal(mu=0.0, sigma=1.0))
x = dp.sample(5)
y = dp.sample(5)
x_data, y_data, theta = sess.run([x, y, dp.theta])
for sample in x_data:
self.assertTrue(sample in theta)
for sample in y_data:
self.assertTrue(sample in theta)
def test_persistent_state(self):
with self.test_session() as sess:
dp = DirichletProcess(0.1, Normal(loc=0.0, scale=1.0))
x = dp.sample(5)
y = dp.sample(5)
x_data, y_data, locs = sess.run([x, y, dp.locs])
for sample in x_data:
self.assertTrue(sample in locs)
for sample in y_data:
self.assertTrue(sample in locs)
# The number of sticks broken is dynamic, changing across evaluations.
sess = tf.Session()
print(sess.run(dp))
print(sess.run(dp))
# Demo of the DirichletProcess random variable in Edward.
# It is associated to a sample tensor, which in turn is associated to
# one of its atoms (base distributions).
base_cls = Normal
kwargs = {'mu': 0.0, 'sigma': 1.0}
# Highly concentrated DP.
alpha = 1.0
dp = DirichletProcess(alpha, base_cls, **kwargs)
x = dp.sample(1000)
samples = sess.run(x)
plt.hist(samples, bins=100, range=(-3.0, 3.0))
plt.title("DP({0}, N(0, 1))".format(alpha))
plt.show()
# More spread out DP.
alpha = 50.0
dp = DirichletProcess(alpha, base_cls, **kwargs)
x = dp.sample(1000)
samples = sess.run(x)
plt.hist(samples, bins=100, range=(-3.0, 3.0))
plt.title("DP({0}, N(0, 1))".format(alpha))
plt.show()
# ``theta`` is the distribution indirectly returned by the DP.
def _test(n, alpha, base_cls, *args, **kwargs): x = DirichletProcess(alpha=alpha, base_cls=base_cls, *args, **kwargs) base = base_cls(*args, **kwargs) val_est = get_dims(x.sample(n)) val_true = n + get_dims(alpha) + get_dims(base) assert val_est == val_true
def _test(self, n, concentration, base):
x = DirichletProcess(concentration=concentration, base=base)
val_est = x.sample(n).shape.as_list()
val_true = n + tf.convert_to_tensor(concentration).shape.as_list() + \
tf.convert_to_tensor(base).shape.as_list()
self.assertEqual(val_est, val_true)
def _test(self, n, alpha, base):
x = DirichletProcess(alpha=alpha, base=base)
val_est = x.sample(n).shape.as_list()
val_true = n + tf.convert_to_tensor(alpha).shape.as_list() + \
tf.convert_to_tensor(base).shape.as_list()
self.assertEqual(val_est, val_true)
def _test(self, n, concentration, base):
x = DirichletProcess(concentration=concentration, base=base)
val_est = x.sample(n).shape.as_list()
val_true = n + tf.convert_to_tensor(concentration).shape.as_list() + \
tf.convert_to_tensor(base).shape.as_list()
self.assertEqual(val_est, val_true)
dp = dirichlet_process(10.0)
# The number of sticks broken is dynamic, changing across evaluations.
sess = tf.Session()
print(sess.run(dp))
print(sess.run(dp))
# Demo of the DirichletProcess random variable in Edward.
base = Normal(0.0, 1.0)
# Highly concentrated DP.
alpha = 1.0
dp = DirichletProcess(alpha, base)
x = dp.sample(1000)
samples = sess.run(x)
plt.hist(samples, bins=100, range=(-3.0, 3.0))
plt.title("DP({0}, N(0, 1))".format(alpha))
plt.show()
# More spread out DP.
alpha = 50.0
dp = DirichletProcess(alpha, base)
x = dp.sample(1000)
samples = sess.run(x)
plt.hist(samples, bins=100, range=(-3.0, 3.0))
plt.title("DP({0}, N(0, 1))".format(alpha))
plt.show()
# States persist across calls to sample() in a DP.