def create_vrnn(generative_class=base.ConditionalNormalDistribution,
batch_size=2, data_size=3, rnn_hidden_size=4,
latent_size=5, fcnet_hidden_size=7, encoded_data_size=9,
encoded_latent_size=11, num_timesteps=7, data_lengths=(7, 4),
use_tilt=False, random_seed=None):
"""Creates a VRNN and some dummy data to feed it for testing purposes.
Args:
generative_class: The class of the generative distribution.
batch_size: The number of elements per batch.
data_size: The dimension of the vectors that make up the data sequences.
rnn_hidden_size: The hidden state dimension of the RNN that forms the
deterministic part of this VRNN.
latent_size: The size of the stochastic latent state of the VRNN.
fcnet_hidden_size: The size of the hidden layer of the fully connected
networks that parameterize the conditional probability distributions
of the VRNN.
encoded_data_size: The size of the output of the data encoding network.
encoded_latent_size: The size of the output of the latent state encoding
network.
num_timesteps: The maximum number of timesteps in the data.
data_lengths: A tuple of size batch_size that contains the desired lengths
of each sequence in the dummy data.
use_tilt: Use a tilting function.
random_seed: A random seed to feed the VRNN, mainly useful for testing
purposes.
Returns:
model: A VRNN object.
inputs: A Tensor of shape [num_timesteps, batch_size, data_size], the inputs
to the model, also known as the observations.
targets: A Tensor of shape [num_timesteps, batch_size, data_size], the
desired outputs of the model.
lengths: A Tensor of shape [batch_size], the lengths of the sequences in the
batch.
"""
fcnet_hidden_sizes = [fcnet_hidden_size]
initializers = {"w": tf.contrib.layers.xavier_initializer(seed=random_seed),
"b": tf.zeros_initializer()}
model = vrnn.create_vrnn(
data_size,
latent_size,
generative_class,
rnn_hidden_size=rnn_hidden_size,
fcnet_hidden_sizes=fcnet_hidden_sizes,
encoded_data_size=encoded_data_size,
encoded_latent_size=encoded_latent_size,
use_tilt=use_tilt,
initializers=initializers,
random_seed=random_seed)
inputs = tf.random_uniform([num_timesteps, batch_size, data_size],
seed=random_seed, dtype=tf.float32)
targets = tf.random_uniform([num_timesteps, batch_size, data_size],
seed=random_seed, dtype=tf.float32)
lengths = tf.constant(data_lengths, dtype=tf.int32)
return model, inputs, targets, lengths
def create_vrnn(generative_class=base.ConditionalNormalDistribution,
batch_size=2, data_size=3, rnn_hidden_size=4,
latent_size=5, fcnet_hidden_size=7, encoded_data_size=9,
encoded_latent_size=11, num_timesteps=7, data_lengths=(7, 4),
use_tilt=False, random_seed=None):
"""Creates a VRNN and some dummy data to feed it for testing purposes.
Args:
generative_class: The class of the generative distribution.
batch_size: The number of elements per batch.
data_size: The dimension of the vectors that make up the data sequences.
rnn_hidden_size: The hidden state dimension of the RNN that forms the
deterministic part of this VRNN.
latent_size: The size of the stochastic latent state of the VRNN.
fcnet_hidden_size: The size of the hidden layer of the fully connected
networks that parameterize the conditional probability distributions
of the VRNN.
encoded_data_size: The size of the output of the data encoding network.
encoded_latent_size: The size of the output of the latent state encoding
network.
num_timesteps: The maximum number of timesteps in the data.
data_lengths: A tuple of size batch_size that contains the desired lengths
of each sequence in the dummy data.
use_tilt: Use a tilting function.
random_seed: A random seed to feed the VRNN, mainly useful for testing
purposes.
Returns:
model: A VRNN object.
inputs: A Tensor of shape [num_timesteps, batch_size, data_size], the inputs
to the model, also known as the observations.
targets: A Tensor of shape [num_timesteps, batch_size, data_size], the
desired outputs of the model.
lengths: A Tensor of shape [batch_size], the lengths of the sequences in the
batch.
"""
fcnet_hidden_sizes = [fcnet_hidden_size]
initializers = {"w": tf.contrib.layers.xavier_initializer(seed=random_seed),
"b": tf.zeros_initializer()}
model = vrnn.create_vrnn(
data_size,
latent_size,
generative_class,
rnn_hidden_size=rnn_hidden_size,
fcnet_hidden_sizes=fcnet_hidden_sizes,
encoded_data_size=encoded_data_size,
encoded_latent_size=encoded_latent_size,
use_tilt=use_tilt,
initializers=initializers,
random_seed=random_seed)
inputs = tf.random_uniform([num_timesteps, batch_size, data_size],
seed=random_seed, dtype=tf.float32)
targets = tf.random_uniform([num_timesteps, batch_size, data_size],
seed=random_seed, dtype=tf.float32)
lengths = tf.constant(data_lengths, dtype=tf.int32)
return model, inputs, targets, lengths
def dummmy_dataset_and_model_fn(self, *unused_args, **unused_kwargs):
# We ignore the arguments in the dummy but need to preserve prototype.
batch_elements = 5
sequence_length = 4
data_dimensions = 3
dataset = tf.data.Dataset.from_tensors(
tf.zeros((sequence_length, batch_elements, data_dimensions),
dtype=tf.float32))
inputs = dataset.make_one_shot_iterator().get_next()
targets = tf.zeros_like(inputs)
lengths = tf.constant([sequence_length] * batch_elements)
mean = tf.constant((0.0, 0.0, 0.0))
model = vrnn.create_vrnn(data_dimensions, 1,
base.ConditionalNormalDistribution)
return inputs, targets, lengths, model, mean
def create_dataset_and_model(config, split, shuffle, repeat):
"""Creates the dataset and model for a given config.
Args:
config: A configuration object with config values accessible as properties.
Most likely a FLAGS object. This function expects the properties
batch_size, dataset_path, dataset_type, and latent_size to be defined.
split: The dataset split to load.
shuffle: If true, shuffle the dataset randomly.
repeat: If true, repeat the dataset endlessly.
Returns:
inputs: A batch of input sequences represented as a dense Tensor of shape
[time, batch_size, data_dimension].
targets: A batch of target sequences represented as a dense Tensor of
shape [time, batch_size, data_dimension].
lens: An int Tensor of shape [batch_size] representing the lengths of each
sequence in the batch.
model: A vrnn.VRNNCell model object.
Raises:
ValueError: if the config is invalid.
"""
sigma_min = 0.0
if config.dataset_type == "pianoroll":
inputs, targets, lengths, mean = datasets.create_pianoroll_dataset(
config.dataset_path,
split,
config.batch_size,
shuffle=shuffle,
repeat=repeat)
# Convert the mean of the training set to logit space so it can be used to
# initialize the bias of the generative distribution.
emission_bias_init = -tf.log(
1. / tf.clip_by_value(mean, 0.0001, 0.9999) - 1)
emission_distribution_class = base.ConditionalBernoulliDistribution
elif config.dataset_type == "speech":
inputs, targets, lengths = datasets.create_speech_dataset(
config.dataset_path,
config.batch_size,
samples_per_timestep=config.data_dimension,
prefetch_buffer_size=1,
shuffle=False,
repeat=False)
# There is no bias for the generative distribution because the test set
# is assumed to be already standardized with the training set statistics.
mean = None
emission_bias_init = None
emission_distribution_class = base.ConditionalNormalDistribution
if config.model == "vrnn":
model = vrnn.create_vrnn(inputs.get_shape().as_list()[2],
config.latent_size,
emission_distribution_class,
emission_bias_init=emission_bias_init,
proposal_type=config.proposal_type,
sigma_min=sigma_min,
raw_sigma_bias=0.5,
use_tilt=(config.bound == "fivo-aux"))
elif config.model == "srnn":
model = srnn.create_srnn(inputs.get_shape().as_list()[2],
config.latent_size,
emission_distribution_class,
emission_bias_init=emission_bias_init,
proposal_type=config.proposal_type,
sigma_min=sigma_min,
raw_sigma_bias=0.5,
use_tilt=(config.bound == "fivo-aux"))
else:
raise ValueError("model flag: %s is unrecognized" % config.model)
return inputs, targets, lengths, model, mean