def create_graph():
"""Creates the dataset, model, and bound."""
xs, lengths = datasets.create_chain_graph_dataset(
config.batch_size,
config.num_timesteps,
steps_per_observation=1,
state_size=1,
transition_variance=config.variance,
observation_variance=config.variance)
model = ghmm.TrainableGaussianHMM(config.num_timesteps,
config.proposal_type,
transition_variances=config.variance,
emission_variances=config.variance,
random_seed=config.random_seed)
true_likelihood = tf.reduce_mean(
model.likelihood(tf.squeeze(xs)) / tf.to_float(lengths))
outs = [true_likelihood]
outs.extend(list(create_bound(model, xs, lengths)))
return outs
def create_graph():
"""Creates the training graph."""
global_step = tf.train.get_or_create_global_step()
xs, lengths = datasets.create_chain_graph_dataset(
config.batch_size,
config.num_timesteps,
steps_per_observation=1,
state_size=1,
transition_variance=config.variance,
observation_variance=config.variance)
model = ghmm.TrainableGaussianHMM(config.num_timesteps,
config.proposal_type,
transition_variances=config.variance,
emission_variances=config.variance,
random_seed=config.random_seed)
loss, bound, true_ll, gap = create_losses(model, xs, lengths)
opt = tf.train.AdamOptimizer(config.learning_rate)
grads = opt.compute_gradients(loss, var_list=tf.trainable_variables())
train_op = opt.apply_gradients(grads, global_step=global_step)
return bound, true_ll, gap, train_op, global_step