def __call__(self, input_layer, num_units, bias=True, name=PROVIDED, stddev=None, init=None, lengths=None):
"""Creates an unrolled GRU to process sequence data.
The initial state is drawn from the bookkeeper's recurrent state and if it
supports state saving, then it is saved.
Args:
input_layer: PrettyTensor (provided).
num_units: Number of units in the hidden states.
bias: Whether or not to use a bias.
name: The name of this layer.
stddev: Standard deviation for Gaussian initialization of parameters.
init: A tf.*Initializer that is used to initialize the variables.
lengths: An optional Tensor that encodes a length for each item in the
minibatch. This is used to truncate computation.
Returns:
A sequence with the result at each timestep.
Raises:
ValueError: if head is not a sequence, the shape is not rank 2 or the
number of nodes (second dim) is not known.
"""
if not self.template:
gru_template = prettytensor.template("input", input_layer.bookkeeper)
self.template = gru_template.gru_cell(
prettytensor.UnboundVariable("state"), num_units, bias=bias, stddev=stddev, init=init
).as_fn("input", "state")
batch_size = input_layer.shape[0]
return unroll_state_saver(input_layer, name, [(batch_size, num_units)], self.template, lengths)
def __call__(self,
input_layer,
num_units,
bias=True,
peephole=True,
name=PROVIDED,
stddev=None,
init=None,
lengths=None):
"""Creates an unrolled LSTM to process sequence data.
The initial state is drawn from the bookkeeper's recurrent state and if it
supports state saving, then it is saved.
Args:
input_layer: PrettyTensor (provided).
num_units: Number of nodes in the hidden states and the output size.
bias: Whether or not to use a bias.
peephole: Whether to use peephole connections.
name: The name of this layer.
stddev: Standard deviation for Gaussian initialization of parameters.
init: A tf.*Initializer that is used to initialize the variables.
lengths: An optional Tensor that encodes a length for each item in the
minibatch. This is used to truncate computation.
Returns:
A sequence with the result at each timestep.
Raises:
ValueError: if head is not a sequence, the shape is not rank 2 or the
number of nodes (second dim) is not known.
"""
if not self.template:
lstm_template = prettytensor.template('input',
input_layer.bookkeeper)
names = ['c', 'h']
states = [prettytensor.UnboundVariable(n) for n in names]
# TODO(eiderman): Move this to pt.make_template() when it is ready.
self.template = lstm_template.lstm_cell(states=states,
num_units=num_units,
bias=bias,
peephole=peephole,
stddev=stddev,
init=init).as_fn(
'input', *names)
batch_size = input_layer.shape[0]
state_shapes = [[batch_size, num_units], [batch_size, num_units]]
return unroll_state_saver(input_layer, name, state_shapes,
self.template, lengths)
def __call__(self,
input_layer,
num_units,
bias=True,
peephole=True,
name=PROVIDED,
stddev=None,
init=None,
lengths=None):
"""Creates an unrolled LSTM to process sequence data.
The initial state is drawn from the bookkeeper's recurrent state and if it
supports state saving, then it is saved.
Args:
input_layer: PrettyTensor (provided).
num_units: Number of nodes in the hidden states and the output size.
bias: Whether or not to use a bias.
peephole: Whether to use peephole connections.
name: The name of this layer.
stddev: Standard deviation for Gaussian initialization of parameters.
init: A tf.*Initializer that is used to initialize the variables.
lengths: An optional Tensor that encodes a length for each item in the
minibatch. This is used to truncate computation.
Returns:
A sequence with the result at each timestep.
Raises:
ValueError: if head is not a sequence, the shape is not rank 2 or the
number of nodes (second dim) is not known.
"""
if not self.template:
lstm_template = prettytensor.template('input', input_layer.bookkeeper)
names = ['c', 'h']
states = [prettytensor.UnboundVariable(n) for n in names]
# TODO(eiderman): Move this to pt.make_template() when it is ready.
self.template = lstm_template.lstm_cell(states=states,
num_units=num_units,
bias=bias,
peephole=peephole,
stddev=stddev,
init=init).as_fn('input', *names)
batch_size = input_layer.shape[0]
state_shapes = [[batch_size, num_units],
[batch_size, num_units]]
return unroll_state_saver(input_layer, name, state_shapes, self.template,
lengths)
def __call__(self,
input_layer,
*args,
**kwargs):
name = kwargs.pop('name', 'sequence')
lengths = kwargs.pop('lengths', None)
state_tuples = state_fn(input_layer, *args, **kwargs)
if not self.template:
template = prettytensor.template('input', input_layer.bookkeeper)
states = [prettytensor.UnboundVariable(s[0]) for s in state_tuples]
self.template = cell_fn(template, states, *args, **kwargs).as_fn(
'input', *[s[0] for s in state_tuples])
batch_size = input_layer.shape[0]
state_shapes = [[batch_size, s[1]] for s in state_tuples]
return unroll_state_saver(
input_layer, name, state_shapes, self.template, lengths)
def __call__(self, input_layer, *args, **kwargs):
name = kwargs.pop('name', 'sequence')
lengths = kwargs.pop('lengths', None)
state_tuples = state_fn(input_layer, *args, **kwargs)
if not self.template:
template = prettytensor.template('input',
input_layer.bookkeeper)
states = [
prettytensor.UnboundVariable(s[0])
for s in state_tuples
]
self.template = cell_fn(template, states, *args,
**kwargs).as_fn(
'input',
*[s[0] for s in state_tuples])
batch_size = input_layer.shape[0]
state_shapes = [[batch_size, s[1]] for s in state_tuples]
return unroll_state_saver(input_layer, name, state_shapes,
self.template, lengths)
def __call__(self,
input_layer,
num_units,
bias=True,
name=PROVIDED,
stddev=None,
init=None,
lengths=None):
"""Creates an unrolled GRU to process sequence data.
The initial state is drawn from the bookkeeper's recurrent state and if it
supports state saving, then it is saved.
Args:
input_layer: PrettyTensor (provided).
num_units: Number of units in the hidden states.
bias: Whether or not to use a bias.
name: The name of this layer.
stddev: Standard deviation for Gaussian initialization of parameters.
init: A tf.*Initializer that is used to initialize the variables.
lengths: An optional Tensor that encodes a length for each item in the
minibatch. This is used to truncate computation.
Returns:
A sequence with the result at each timestep.
Raises:
ValueError: if head is not a sequence, the shape is not rank 2 or the
number of nodes (second dim) is not known.
"""
if not self.template:
gru_template = prettytensor.template('input',
input_layer.bookkeeper)
self.template = gru_template.gru_cell(
prettytensor.UnboundVariable('state'),
num_units,
bias=bias,
stddev=stddev,
init=init).as_fn('input', 'state')
batch_size = input_layer.shape[0]
return unroll_state_saver(input_layer, name, [(batch_size, num_units)],
self.template, lengths)
