def test_all_finite_raises(self):
with self.test_session():
x = np.inf * tf.constant([-1.0, -2.0, -3.0, -4.0])
with self.assertRaisesOpError('Inf'):
log_mean_exp(x).eval()
x = tf.constant([-1.0, np.nan, -3.0, -4.0])
with self.assertRaisesOpError('NaN'):
log_mean_exp(x).eval()
def test_all_finite_raises(self):
with self.test_session():
x = np.inf * tf.constant([-1.0, -2.0, -3.0, -4.0])
with self.assertRaisesOpError('Inf'):
log_mean_exp(x).eval()
x = tf.constant([-1.0, np.nan, -3.0, -4.0])
with self.assertRaisesOpError('NaN'):
log_mean_exp(x).eval()
def test_log_mean_exp_2d(self):
with self.test_session():
x = tf.constant([[-1.0], [-2.0], [-3.0], [-4.0]])
self.assertAllClose(log_mean_exp(x).eval(), -1.9461046625586951)
x = tf.constant([[-1.0, -2.0], [-3.0, -4.0]])
self.assertAllClose(log_mean_exp(x).eval(), -1.9461046625586951)
self.assertAllClose(
log_mean_exp(x, 0).eval(),
np.array([-1.5662191695169727, -2.5662191695169727]))
self.assertAllClose(
log_mean_exp(x, 1).eval(),
np.array([-1.3798854930417224, -3.3798854930417224]))
def test_log_mean_exp_2d(self):
with self.test_session():
x = tf.constant([[-1.0], [-2.0], [-3.0], [-4.0]])
self.assertAllClose(log_mean_exp(x).eval(),
-1.9461046625586951)
x = tf.constant([[-1.0, -2.0], [-3.0, -4.0]])
self.assertAllClose(log_mean_exp(x).eval(),
-1.9461046625586951)
self.assertAllClose(log_mean_exp(x, 0).eval(),
np.array([-1.5662191695169727,
-2.5662191695169727]))
self.assertAllClose(log_mean_exp(x, 1).eval(),
np.array([-1.3798854930417224,
-3.3798854930417224]))
def build_reparam_loss(self):
"""Build loss function. Its automatic differentiation
is a stochastic gradient of
.. math::
-E_{q(z^1; \lambda), ..., q(z^K; \lambda)} [
\log 1/K \sum_{k=1}^K p(x, z^k)/q(z^k; \lambda) ]
based on the reparameterization trick. (Kingma and Welling, 2014)
Computed by sampling from :math:`q(z;\lambda)` and evaluating
the expectation using Monte Carlo sampling. Note there is a
difference between the number of samples to approximate the
expectations (`n_samples`) and the number of importance
samples to determine how many expectations (`K`).
"""
x = self.data
for s in range(self.n_samples):
z = self.variational.sample(self.K)
p_log_prob = self.model.log_prob(x, z)
q_log_prob = self.variational.log_prob(z)
log_w = p_log_prob - q_log_prob
losses += [log_mean_exp(log_w)]
losses = tf.pack(losses)
self.loss = tf.reduce_mean(losses)
return -self.loss
def build_reparam_loss(self):
"""Build loss function. Its automatic differentiation
is a stochastic gradient of
.. math::
-E_{q(z^1; \lambda), ..., q(z^K; \lambda)} [
\log 1/K \sum_{k=1}^K p(x, z^k)/q(z^k; \lambda) ]
based on the reparameterization trick. (Kingma and Welling, 2014)
Computed by sampling from :math:`q(z;\lambda)` and evaluating
the expectation using Monte Carlo sampling. Note there is a
difference between the number of samples to approximate the
expectations (`n_minibatch`) and the number of importance
samples to determine how many expectations (`K`).
"""
x = self.data
for s in range(self.n_minibatch):
z = self.variational.sample(self.K)
p_log_prob = self.model.log_prob(x, z)
q_log_prob = self.variational.log_prob(z)
log_w = p_log_prob - q_log_prob
losses += [log_mean_exp(log_w)]
losses = tf.pack(losses)
self.loss = tf.reduce_mean(losses)
return -self.loss
def build_score_loss(self):
"""Build loss function. Its automatic differentiation
is a stochastic gradient of
.. math::
-E_{q(z^1; \lambda), ..., q(z^K; \lambda)} [
\log 1/K \sum_{k=1}^K p(x, z^k)/q(z^k; \lambda) ]
based on the score function estimator. (Paisley et al., 2012)
Computed by sampling from :math:`q(z;\lambda)` and evaluating
the expectation using Monte Carlo sampling. Note there is a
difference between the number of samples to approximate the
expectations (`n_samples`) and the number of importance
samples to determine how many expectations (`K`).
"""
x = self.data
losses = []
for s in range(self.n_samples):
z = self.variational.sample(self.K)
p_log_prob = self.model.log_prob(x, z)
q_log_prob = self.variational.log_prob(stop_gradient(z))
log_w = p_log_prob - q_log_prob
losses += [log_mean_exp(log_w)]
losses = tf.pack(losses)
self.loss = tf.reduce_mean(losses)
return -tf.reduce_mean(q_log_prob * stop_gradient(losses))
def build_loss_and_gradients(self, var_list):
"""Build loss function. Its automatic differentiation
is a stochastic gradient of
.. math::
-E_{q(z^1; \lambda), ..., q(z^K; \lambda)} [
\log 1/K \sum_{k=1}^K p(x, z^k)/q(z^k; \lambda) ]
based on the score function estimator. (Paisley et al., 2012)
Computed by sampling from :math:`q(z;\lambda)` and evaluating
the expectation using Monte Carlo sampling. Note there is a
difference between the number of samples to approximate the
expectations (`n_samples`) and the number of importance
samples to determine how many expectations (`K`).
"""
x = self.data
# Form n_samples x K matrix of log importance weights.
log_w = []
for s in range(self.n_samples * self.K):
z_sample = {}
q_log_prob = 0.0
for z, qz in six.iteritems(self.latent_vars):
# Copy q(z) to obtain new set of posterior samples.
qz_copy = copy(qz, scope='inference_' + str(s))
z_sample[z] = qz_copy.value()
q_log_prob += tf.reduce_sum(
qz.log_prob(tf.stop_gradient(z_sample[z])))
p_log_prob = self.model_wrapper.log_prob(x, z_sample)
log_w += [p_log_prob - q_log_prob]
log_w = tf.reshape(log_w, [self.n_samples, self.K])
# Take log mean exp across importance weights (columns).
losses = log_mean_exp(log_w, 1)
loss = -tf.reduce_mean(losses)
if var_list is None:
var_list = tf.trainable_variables()
grads = tf.gradients(
-tf.reduce_mean(q_log_prob * tf.stop_gradient(losses)),
[v.ref() for v in var_list])
grads_and_vars = list(zip(grads, var_list))
return loss, grads_and_vars
def build_loss_and_gradients(self, var_list):
"""Build loss function. Its automatic differentiation
is a stochastic gradient of
.. math::
-E_{q(z^1; \lambda), ..., q(z^K; \lambda)} [
\log 1/K \sum_{k=1}^K p(x, z^k)/q(z^k; \lambda) ]
based on the score function estimator. (Paisley et al., 2012)
Computed by sampling from :math:`q(z;\lambda)` and evaluating
the expectation using Monte Carlo sampling. Note there is a
difference between the number of samples to approximate the
expectations (`n_samples`) and the number of importance
samples to determine how many expectations (`K`).
"""
x = self.data
# Form n_samples x K matrix of log importance weights.
log_w = []
for s in range(self.n_samples * self.K):
z_sample = {}
q_log_prob = 0.0
for z, qz in six.iteritems(self.latent_vars):
# Copy q(z) to obtain new set of posterior samples.
qz_copy = copy(qz, scope='inference_' + str(s))
z_sample[z] = qz_copy.value()
q_log_prob += tf.reduce_sum(qz.log_prob(tf.stop_gradient(z_sample[z])))
p_log_prob = self.model_wrapper.log_prob(x, z_sample)
log_w += [p_log_prob - q_log_prob]
log_w = tf.reshape(log_w, [self.n_samples, self.K])
# Take log mean exp across importance weights (columns).
losses = log_mean_exp(log_w, 1)
loss = -tf.reduce_mean(losses)
if var_list is None:
var_list = tf.trainable_variables()
grads = tf.gradients(
-tf.reduce_mean(q_log_prob * tf.stop_gradient(losses)),
[v.ref() for v in var_list])
grads_and_vars = list(zip(grads, var_list))
return loss, grads_and_vars
def build_loss_and_gradients(self, var_list):
"""Build loss function. Its automatic differentiation
is a stochastic gradient of
.. math::
-E_{q(z^1; \lambda), ..., q(z^K; \lambda)} [
\log 1/K \sum_{k=1}^K p(x, z^k)/q(z^k; \lambda) ]
based on the reparameterization trick.
"""
# Form vector of K log importance weights.
log_w = []
for k in range(self.K):
scope = 'inference_' + str(id(self)) + '/' + str(k)
z_sample = {}
q_log_prob = 0.0
for z, qz in six.iteritems(self.latent_vars):
# Copy q(z) to obtain new set of posterior samples.
qz_copy = copy(qz, scope=scope)
z_sample[z] = qz_copy
q_log_prob += tf.reduce_sum(qz_copy.log_prob(qz_copy))
p_log_prob = 0.0
for z in six.iterkeys(self.latent_vars):
# Copy p(z), swapping its conditioning set with samples
# from variational distribution.
z_copy = copy(z, z_sample, scope=scope)
p_log_prob += tf.reduce_sum(z_copy.log_prob(z_sample[z]))
for x, qx in six.iteritems(self.data):
if isinstance(x, RandomVariable):
# Copy p(x | z), swapping its conditioning set with samples
# from variational distribution.
x_copy = copy(x, z_sample, scope=scope)
p_log_prob += tf.reduce_sum(x_copy.log_prob(qx))
log_w += [p_log_prob - q_log_prob]
loss = -log_mean_exp(log_w)
grads = tf.gradients(loss, [v.ref() for v in var_list])
grads_and_vars = list(zip(grads, var_list))
return loss, grads_and_vars
def build_reparam_loss(self):
"""Build loss function. Its automatic differentiation
is a stochastic gradient of
.. math::
-E_{q(z^1; \lambda), ..., q(z^K; \lambda)} [
\log 1/K \sum_{k=1}^K p(x, z^k)/q(z^k; \lambda) ]
based on the reparameterization trick. (Kingma and Welling, 2014)
Computed by sampling from :math:`q(z;\lambda)` and evaluating
the expectation using Monte Carlo sampling. Note there is a
difference between the number of samples to approximate the
expectations (`n_samples`) and the number of importance
samples to determine how many expectations (`K`).
"""
x = self.data
# Form n_samples x K matrix of log importance weights.
log_w = []
for s in range(self.n_samples * self.K):
z_sample = {}
q_log_prob = 0.0
for z, qz in six.iteritems(self.latent_vars):
# Copy q(z) to obtain new set of posterior samples.
qz_copy = copy(qz, scope='inference_' + str(s))
z_sample[z] = qz_copy.value()
q_log_prob += tf.reduce_sum(qz.log_prob(z_sample[z]))
p_log_prob = self.model_wrapper.log_prob(x, z_sample)
log_w += [p_log_prob - q_log_prob]
log_w = tf.reshape(log_w, [self.n_samples, self.K])
# Take log mean exp across importance weights (columns).
losses = log_mean_exp(log_w, 1)
self.loss = tf.reduce_mean(losses)
return -self.loss
def test_log_mean_exp_1d(self):
with self.test_session():
x = tf.constant([-1.0, -2.0, -3.0, -4.0])
self.assertAllClose(log_mean_exp(x).eval(),
-1.9461046625586951)
def test_2d():
x = tf.constant([[-1.0], [-2.0], [-3.0], [-4.0]])
val_ed = log_mean_exp(x)
val_true = -1.9461046625586951
assert np.allclose(val_ed.eval(), val_true)
def test_log_mean_exp_2d(self):
with self.test_session():
x = tf.constant([[-1.0], [-2.0], [-3.0], [-4.0]])
self.assertAllClose(log_mean_exp(x).eval(), -1.9461046625586951)
