def create_bound(model, xs, lengths):
"""Creates the bound to be evaluated."""
if config.bound == "elbo":
ll_per_seq, log_weights, _ = bounds.iwae(
model,
xs,
lengths,
num_samples=1,
parallel_iterations=config.parallel_iterations)
elif config.bound == "iwae":
ll_per_seq, log_weights, _ = bounds.iwae(
model,
xs,
lengths,
num_samples=config.num_samples,
parallel_iterations=config.parallel_iterations)
elif config.bound == "fivo":
ll_per_seq, log_weights, resampled, _ = bounds.fivo(
model,
xs,
lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations)
# Compute bound scaled by number of timesteps.
bound_per_t = ll_per_seq / tf.to_float(lengths)
if config.bound == "fivo":
return bound_per_t, log_weights, resampled
else:
return bound_per_t, log_weights
def create_bound(model, xs, lengths):
"""Creates the bound to be evaluated."""
if config.bound == "elbo":
ll_per_seq, log_weights, _ = bounds.iwae(
model, xs, lengths, num_samples=1,
parallel_iterations=config.parallel_iterations
)
elif config.bound == "iwae":
ll_per_seq, log_weights, _ = bounds.iwae(
model, xs, lengths, num_samples=config.num_samples,
parallel_iterations=config.parallel_iterations
)
elif config.bound == "fivo":
ll_per_seq, log_weights, resampled, _ = bounds.fivo(
model, xs, lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations
)
# Compute bound scaled by number of timesteps.
bound_per_t = ll_per_seq / tf.to_float(lengths)
if config.bound == "fivo":
return bound_per_t, log_weights, resampled
else:
return bound_per_t, log_weights
def create_loss():
"""Creates the loss to be optimized.
Returns:
bound: A float Tensor containing the value of the bound that is
being optimized.
loss: A float Tensor that when differentiated yields the gradients
to apply to the model. Should be optimized via gradient descent.
"""
inputs, targets, lengths, model, _ = create_dataset_and_model_fn(
config, split="train", shuffle=True, repeat=True)
# Compute lower bounds on the log likelihood
if config.bound == "elbo":
ll_per_seq, _, _ = bounds.iwae(
model, (inputs, targets),
lengths,
num_samples=1,
parallel_iterations=config.parallel_iterations)
elif config.bound == "iwae":
ll_per_seq, _, _ = bounds.iwae(
model, (inputs, targets),
lengths,
num_samples=config.num_samples,
parallel_iterations=config.parallel_iterations)
elif config.bound in ("fivo", "fivo-aux"):
if config.resampling_type == "relaxed":
ll_per_seq, _, _, _ = bounds.fivo(
model, (inputs, targets),
lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
random_seed=config.random_seed,
relaxed_resampling_temperature=config.
relaxed_resampling_temperature,
parallel_iterations=config.parallel_iterations)
else:
ll_per_seq, _, _, _ = bounds.fivo(
model, (inputs, targets),
lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations)
# Compute loss scaled by number of timesteps
ll_per_t = tf.reduce_mean(ll_per_seq / tf.to_float(lengths))
ll_per_seq = tf.reduce_mean(ll_per_seq)
tf.summary.scalar("train_ll_per_seq", ll_per_seq)
tf.summary.scalar("train_ll_per_t", ll_per_t)
if config.normalize_by_seq_len:
return ll_per_t, -ll_per_t
else:
return ll_per_seq, -ll_per_seq
def create_losses(model, observations, lengths):
"""Creates the loss to be optimized.
Args:
model: A Trainable GHMM model.
observations: A set of observations.
lengths: The lengths of each sequence in the observations.
Returns:
loss: A float Tensor that when differentiated yields the gradients
to apply to the model. Should be optimized via gradient descent.
bound: A float Tensor containing the value of the bound that is
being optimized.
true_ll: The true log-likelihood of the data under the model.
bound_gap: The gap between the bound and the true log-likelihood.
"""
# Compute lower bounds on the log likelihood.
if config.bound == "elbo":
ll_per_seq, _, _ = bounds.iwae(
model, observations, lengths, num_samples=1,
parallel_iterations=config.parallel_iterations
)
elif config.bound == "iwae":
ll_per_seq, _, _ = bounds.iwae(
model, observations, lengths, num_samples=config.num_samples,
parallel_iterations=config.parallel_iterations
)
elif config.bound == "fivo":
if config.resampling_type == "relaxed":
ll_per_seq, _, _, _ = bounds.fivo(
model,
observations,
lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
relaxed_resampling_temperature=config.
relaxed_resampling_temperature,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations)
else:
ll_per_seq, _, _, _ = bounds.fivo(
model, observations, lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations
)
ll_per_t = tf.reduce_mean(ll_per_seq / tf.to_float(lengths))
# Compute the data's true likelihood under the model and the bound gap.
true_ll_per_seq = model.likelihood(tf.squeeze(observations))
true_ll_per_t = tf.reduce_mean(true_ll_per_seq / tf.to_float(lengths))
bound_gap = true_ll_per_seq - ll_per_seq
bound_gap = tf.reduce_mean(bound_gap/ tf.to_float(lengths))
tf.summary.scalar("train_ll_bound", ll_per_t)
tf.summary.scalar("train_true_ll", true_ll_per_t)
tf.summary.scalar("bound_gap", bound_gap)
return -ll_per_t, ll_per_t, true_ll_per_t, bound_gap
def create_losses(model, observations, lengths):
"""Creates the loss to be optimized.
Args:
model: A Trainable GHMM model.
observations: A set of observations.
lengths: The lengths of each sequence in the observations.
Returns:
loss: A float Tensor that when differentiated yields the gradients
to apply to the model. Should be optimized via gradient descent.
bound: A float Tensor containing the value of the bound that is
being optimized.
true_ll: The true log-likelihood of the data under the model.
bound_gap: The gap between the bound and the true log-likelihood.
"""
# Compute lower bounds on the log likelihood.
if config.bound == "elbo":
ll_per_seq, _, _ = bounds.iwae(
model, observations, lengths, num_samples=1,
parallel_iterations=config.parallel_iterations
)
elif config.bound == "iwae":
ll_per_seq, _, _ = bounds.iwae(
model, observations, lengths, num_samples=config.num_samples,
parallel_iterations=config.parallel_iterations
)
elif config.bound == "fivo":
if config.resampling_type == "relaxed":
ll_per_seq, _, _, _ = bounds.fivo(
model,
observations,
lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
relaxed_resampling_temperature=config.
relaxed_resampling_temperature,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations)
else:
ll_per_seq, _, _, _ = bounds.fivo(
model, observations, lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
resampling_type=config.resampling_type,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations
)
ll_per_t = tf.reduce_mean(ll_per_seq / tf.to_float(lengths))
# Compute the data's true likelihood under the model and the bound gap.
true_ll_per_seq = model.likelihood(tf.squeeze(observations))
true_ll_per_t = tf.reduce_mean(true_ll_per_seq / tf.to_float(lengths))
bound_gap = true_ll_per_seq - ll_per_seq
bound_gap = tf.reduce_mean(bound_gap/ tf.to_float(lengths))
tf.summary.scalar("train_ll_bound", ll_per_t)
tf.summary.scalar("train_true_ll", true_ll_per_t)
tf.summary.scalar("bound_gap", bound_gap)
return -ll_per_t, ll_per_t, true_ll_per_t, bound_gap
def test_iwae(self):
"""A golden-value test for the IWAE bound."""
tf.set_random_seed(1234)
with self.test_session() as sess:
model, inputs, targets, lengths = create_vrnn(random_seed=1234)
outs = bounds.iwae(model, (inputs, targets),
lengths,
num_samples=4,
parallel_iterations=1)
sess.run(tf.global_variables_initializer())
log_p_hat, weights, _ = sess.run(outs)
self.assertAllClose([-23.301426, -13.64028], log_p_hat)
weights_gt = np.array(
[[[-3.66708851, -2.07074022, -4.91751671, -5.03293562],
[-2.99690723, -3.17782736, -4.50084877, -3.48536515]],
[[-6.2539978, -4.37615728, -7.43738699, -7.85044909],
[-8.27518654, -6.71545124, -8.96198845, -7.05567837]],
[[-9.19093227, -8.01637268, -11.64603615, -10.51128292],
[-12.34527206, -11.54284477, -11.8667469, -9.69417381]],
[[-12.20609856, -10.47217369, -13.66270638, -13.46115875],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]],
[[-16.14766312, -15.57472229, -17.47755432, -17.98189926],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]],
[[-20.07182884, -18.43191147, -20.1606636, -21.45263863],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]],
[[-24.10270691, -22.20865822, -24.14675522, -25.27248383],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]]])
self.assertAllClose(weights_gt, weights)
def test_iwae(self):
"""A golden-value test for the IWAE bound."""
tf.set_random_seed(1234)
with self.test_session() as sess:
model, inputs, targets, lengths = create_vrnn(random_seed=1234)
outs = bounds.iwae(model, (inputs, targets), lengths, num_samples=4,
parallel_iterations=1)
sess.run(tf.global_variables_initializer())
log_p_hat, weights, _ = sess.run(outs)
self.assertAllClose([-23.301426, -13.64028], log_p_hat)
weights_gt = np.array(
[[[-3.66708851, -2.07074022, -4.91751671, -5.03293562],
[-2.99690723, -3.17782736, -4.50084877, -3.48536515]],
[[-6.2539978, -4.37615728, -7.43738699, -7.85044909],
[-8.27518654, -6.71545124, -8.96198845, -7.05567837]],
[[-9.19093227, -8.01637268, -11.64603615, -10.51128292],
[-12.34527206, -11.54284477, -11.8667469, -9.69417381]],
[[-12.20609856, -10.47217369, -13.66270638, -13.46115875],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]],
[[-16.14766312, -15.57472229, -17.47755432, -17.98189926],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]],
[[-20.07182884, -18.43191147, -20.1606636, -21.45263863],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]],
[[-24.10270691, -22.20865822, -24.14675522, -25.27248383],
[-17.17656708, -16.25190353, -15.28658581, -12.33067703]]])
self.assertAllClose(weights_gt, weights)
def create_graph():
"""Creates the evaluation graph.
Returns:
lower_bounds: A tuple of float Tensors containing the values of the 3
evidence lower bounds, summed across the batch.
total_batch_length: The total number of timesteps in the batch, summed
across batch examples.
batch_size: The batch size.
global_step: The global step the checkpoint was loaded from.
"""
global_step = tf.train.get_or_create_global_step()
inputs, targets, lengths, model, _ = create_dataset_and_model_fn(
config, split=config.split, shuffle=False, repeat=False)
# Compute lower bounds on the log likelihood.
elbo_ll_per_seq, _, _ = bounds.iwae(
model, (inputs, targets),
lengths,
num_samples=1,
parallel_iterations=config.parallel_iterations)
iwae_ll_per_seq, _, _ = bounds.iwae(
model, (inputs, targets),
lengths,
num_samples=config.num_samples,
parallel_iterations=config.parallel_iterations)
# The resampling type should only be used for training, so we ignore it.
fivo_ll_per_seq, _, _, _ = bounds.fivo(
model, (inputs, targets),
lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
random_seed=config.random_seed,
parallel_iterations=config.parallel_iterations)
elbo_ll = tf.reduce_sum(elbo_ll_per_seq)
iwae_ll = tf.reduce_sum(iwae_ll_per_seq)
fivo_ll = tf.reduce_sum(fivo_ll_per_seq)
batch_size = tf.shape(lengths)[0]
total_batch_length = tf.reduce_sum(lengths)
return ((elbo_ll, iwae_ll, fivo_ll), total_batch_length, batch_size,
global_step)
def test_elbo(self):
"""A golden-value test for the ELBO (the IWAE bound with num_samples=1)."""
tf.set_random_seed(1234)
with self.test_session() as sess:
model, inputs, targets, lengths = create_vrnn(random_seed=1234)
outs = bounds.iwae(model, (inputs, targets), lengths, num_samples=1,
parallel_iterations=1)
sess.run(tf.global_variables_initializer())
log_p_hat, _, _ = sess.run(outs)
self.assertAllClose([-21.615765, -13.614225], log_p_hat)
def create_graph():
"""Creates the graph to sample from the model.
First, the model is conditioned on a prefix by sampling a batch of data
and trimming it to prefix_length. The configured bound is used to do the
conditioning. Then the final state from the conditioning is used to sample
from the model.
Returns:
samples: A Tensor of shape [sample_length, batch_size,
num_samples, data_dimension] representing samples from the model.
prefixes: A Tensor of shape [prefix_length, batch_size, data_dimension]
representing the prefixes the model was conditioned on.
"""
inputs, targets, lengths, model, mean = create_dataset_and_model_fn(
config, split=config.split, shuffle=True, repeat=True)
input_prefixes = inputs[:config.prefix_length]
target_prefixes = targets[:config.prefix_length]
prefix_lengths = tf.ones_like(lengths) * config.prefix_length
if config.bound == "elbo":
_, _, state = bounds.iwae(model, (input_prefixes, target_prefixes),
prefix_lengths,
num_samples=1)
elif config.bound == "iwae":
_, _, state = bounds.iwae(model, (input_prefixes, target_prefixes),
prefix_lengths,
num_samples=config.num_samples)
elif config.bound == "fivo":
_, _, _, state = bounds.fivo(
model, (input_prefixes, target_prefixes),
prefix_lengths,
num_samples=config.num_samples,
resampling_criterion=smc.ess_criterion,
random_seed=config.random_seed)
sample_inputs = tf.tile(inputs[config.prefix_length],
[config.num_samples, 1])
samples = sample_from_model(model, state, sample_inputs, mean)
return samples, target_prefixes
def test_elbo(self):
"""A golden-value test for the ELBO (the IWAE bound with num_samples=1)."""
tf.set_random_seed(1234)
with self.test_session() as sess:
model, inputs, targets, lengths = create_vrnn(random_seed=1234)
outs = bounds.iwae(model, (inputs, targets),
lengths,
num_samples=1,
parallel_iterations=1)
sess.run(tf.global_variables_initializer())
log_p_hat, _, _ = sess.run(outs)
self.assertAllClose([-21.615765, -13.614225], log_p_hat)
