def create_cells(self):
"""
Creates a Tensorflow RNN cell object by using the given configuration.
"""
self.cell = get_rnn_cell(scope='rnn_cell',
reuse=self.reuse,
**self.rnn_config)
self.initial_states = self.cell.zero_state(batch_size=self.batch_size,
dtype=tf.float32)
def build_cell(self):
"""
Builds a Tensorflow RNN cell object by using the given configuration `self.rnn_layer_config`.
"""
if self.stack_fw_bw_cells:
single_cell_config = self.rnn_layer_config.copy()
single_cell_config['num_layers'] = 1
for i in range(self.rnn_layer_config['num_layers']):
cell_fw = get_rnn_cell(scope='rnn_cell_fw', reuse=self.reuse, **single_cell_config)
self.cells_fw.append(cell_fw)
self.initial_states_fw.append(cell_fw.zero_state(batch_size=self.batch_size, dtype=tf.float32))
cell_bw = get_rnn_cell(scope='rnn_cell_bw', reuse=self.reuse, **single_cell_config)
self.cells_bw.append(cell_bw)
self.initial_states_bw.append(cell_bw.zero_state(batch_size=self.batch_size, dtype=tf.float32))
else:
cell_fw = get_rnn_cell(scope='rnn_cell_fw', reuse=self.reuse, **self.rnn_layer_config)
self.cells_fw.append(cell_fw)
self.initial_states_fw.append(cell_fw.zero_state(batch_size=self.batch_size, dtype=tf.float32))
cell_bw = get_rnn_cell(scope='rnn_cell_bw', reuse=self.reuse, **self.rnn_layer_config)
self.cells_bw.append(cell_bw)
self.initial_states_bw.append(cell_bw.zero_state(batch_size=self.batch_size, dtype=tf.float32))
def create_cells(self):
if self.stack_fw_bw_cells:
single_cell_config = self.rnn_config.copy()
single_cell_config['num_layers'] = 1
for i in range(self.rnn_config['num_layers']):
cell_fw = get_rnn_cell(scope='rnn_cell_fw',
reuse=self.reuse,
**single_cell_config)
self.cells_fw.append(cell_fw)
self.initial_states_fw.append(
cell_fw.zero_state(batch_size=self.batch_size,
dtype=tf.float32))
cell_bw = get_rnn_cell(scope='rnn_cell_bw',
reuse=self.reuse,
**single_cell_config)
self.cells_bw.append(cell_bw)
self.initial_states_bw.append(
cell_bw.zero_state(batch_size=self.batch_size,
dtype=tf.float32))
else:
cell_fw = get_rnn_cell(scope='rnn_cell_fw',
reuse=self.reuse,
**self.rnn_config)
self.cells_fw.append(cell_fw)
self.initial_states_fw.append(
cell_fw.zero_state(batch_size=self.batch_size,
dtype=tf.float32))
cell_bw = get_rnn_cell(scope='rnn_cell_bw',
reuse=self.reuse,
**self.rnn_config)
self.cells_bw.append(cell_bw)
self.initial_states_bw.append(
cell_bw.zero_state(batch_size=self.batch_size,
dtype=tf.float32))
def __init__(self, reuse, mode, config):
"""
Args:
config (dict): In addition to standard <key, value> pairs, stores the following dictionaries for rnn and
output configurations.
config['output'] = {}
config['output']['keys']
config['output']['dims']
config['output']['activation_funcs']
config['*_rnn'] = {}
config['*_rnn']['num_layers'] (default: 1)
config['*_rnn']['cell_type'] (default: lstm)
config['*_rnn']['size'] (default: 512)
reuse: reuse model parameters.
mode: 'training' or 'sampling'.
"""
self.input_dims = config['input_dims']
self.h_dim = config['latent_hidden_size']
self.z_dim = config['latent_size']
self.additive_q_mu = config['additive_q_mu']
self.dropout_keep_prob = config.get('input_keep_prop', 1)
self.num_linear_layers = config.get('num_fc_layers', 1)
self.use_latent_h_in_outputs = config.get('use_latent_h_in_outputs',
True)
self.use_batch_norm = config['use_batch_norm_fc']
self.reuse = reuse
self.mode = mode
self.is_sampling = mode == 'sampling'
if not (mode == "training"):
self.dropout_keep_prob = 1.0
self.output_config = config['output']
self.output_size_ = [self.z_dim] * 4
self.output_size_.extend(
self.output_config['dims']
) # q_mu, q_sigma, p_mu, p_sigma + model outputs
self.state_size_ = []
# Optional. Linear layers will be used if not passed.
self.input_rnn = False
if 'input_rnn' in config and not (config['input_rnn'] is None) and len(
config['input_rnn'].keys()) > 0:
self.input_rnn = True
self.input_rnn_config = config['input_rnn']
self.input_rnn_cell = get_rnn_cell(scope='input_rnn',
**config['input_rnn'])
self.state_size_.append(self.input_rnn_cell.state_size)
self.latent_rnn_config = config['latent_rnn']
self.latent_rnn_cell_type = config['latent_rnn']['cell_type']
self.latent_rnn_cell = get_rnn_cell(scope='latent_rnn',
**config['latent_rnn'])
self.state_size_.append(self.latent_rnn_cell.state_size)
# Optional. Linear layers will be used if not passed.
self.output_rnn = False
if 'output_rnn' in config and not (
config['output_rnn'] is None) and len(
config['output_rnn'].keys()) > 0:
self.output_rnn = True
self.output_rnn_config = config['output_rnn']
self.output_rnn_cell = get_rnn_cell(scope='output_rnn',
**config['output_rnn'])
self.state_size_.append(self.output_rnn_cell.state_size)
self.activation_func = get_activation_fn(
config.get('fc_layer_activation_func', 'relu'))
self.sigma_func = get_activation_fn('softplus')
def __init__(self, reuse, mode, sample_fn, config):
"""
Args:
reuse: reuse model parameters.
mode: 'training' or 'sampling'.
sample_fn: function to generate sample given model outputs.
config (dict): In addition to standard <key, value> pairs, stores the following dictionaries for rnn and
output configurations.
config['output_layer'] = {}
config['output_layer']['out_keys']
config['output_layer']['out_dims']
config['output_layer']['out_activation_fn']
config['*_rnn'] = {}
config['*_rnn']['num_layers'] (default: 1)
config['*_rnn']['cell_type'] (default: lstm)
config['*_rnn']['size'] (default: 512)
"""
self.reuse = reuse
self.mode = mode
self.sample_fn = sample_fn
self.is_sampling = mode == 'sampling'
self.is_evaluation = mode == "validation" or mode == "test"
self.input_dims = config['input_dims']
self.h_dim = config['hidden_size']
self.latent_h_dim = config.get('latent_hidden_size', self.h_dim)
self.z_dim = config['latent_size']
self.additive_q_mu = config['additive_q_mu']
self.dropout_keep_prob = config.get('input_keep_prop', 1)
self.num_linear_layers = config.get('num_fc_layers', 1)
self.use_latent_h_in_outputs = config.get('use_latent_h_in_outputs',
True)
self.use_batch_norm = config['use_batch_norm_fc']
if not (mode == "training"):
self.dropout_keep_prob = 1.0
self.output_config = config['output_layer']
self.output_size_ = [self.z_dim] * 4
self.output_size_.extend(
self.output_config['out_dims']
) # q_mu, q_sigma, p_mu, p_sigma + model outputs
self.state_size_ = []
# Optional. Linear layers will be used if not passed.
self.input_rnn = False
if 'input_rnn' in config and not (config['input_rnn'] is None) and len(
config['input_rnn'].keys()) > 0:
self.input_rnn = True
self.input_rnn_config = config['input_rnn']
self.input_rnn_cell = get_rnn_cell(scope='input_rnn',
**config['input_rnn'])
# Variational dropout
if config['input_rnn'].get('use_variational_dropout', False):
# TODO input dimensions are hard-coded.
self.input_rnn_cell = tf.contrib.rnn.DropoutWrapper(
self.input_rnn_cell,
input_keep_prob=self.dropout_keep_prob,
output_keep_prob=self.dropout_keep_prob,
variational_recurrent=True,
input_size=(216),
dtype=tf.float32)
self.dropout_keep_prob = 1.0
self.state_size_.append(self.input_rnn_cell.state_size)
self.latent_rnn_config = config['latent_rnn']
self.latent_rnn_cell_type = config['latent_rnn']['cell_type']
self.latent_rnn_cell = get_rnn_cell(scope='latent_rnn',
**config['latent_rnn'])
self.state_size_.append(self.latent_rnn_cell.state_size)
# Optional. Linear layers will be used if not passed.
self.output_rnn = False
if 'output_rnn' in config and not (
config['output_rnn'] is None) and len(
config['output_rnn'].keys()) > 0:
self.output_rnn = True
self.output_rnn_config = config['output_rnn']
self.output_rnn_cell = get_rnn_cell(scope='output_rnn',
**config['output_rnn'])
self.state_size_.append(self.output_rnn_cell.state_size)
self.activation_func = get_activation_fn(
config.get('fc_layer_activation_func', 'relu'))
self.sigma_activaction_fn = tf.nn.softplus
def build_cell(self):
"""
Builds a Tensorflow RNN cell object by using the given configuration `self.rnn_layer_config`.
"""
self.cell = get_rnn_cell(scope='rnn_cell', reuse=self.reuse, **self.rnn_layer_config)
self.initial_states = self.cell.zero_state(batch_size=self.batch_size, dtype=tf.float32)