def delay(x, initial_state=None, time_step=1, name=''):
'''
This function combines :func:`~cntk.ops.sequence.past_value` and :func:`~cntk.ops.sequence.future_value` into a single function.
This is useful when the time_step is computed and can be positive, negative, or 0.
Args:
x: the tensor (or its name) from which the past value is obtained
initial_state: tensor or scalar representing the initial value to be used when the input tensor is shifted in time.
time_step (int): the number of time steps to look into the past, where negative values mean to look into the future, and 0 means a no-op (default 1).
name (str, optional): the name of the Function instance in the network
'''
from ...ops import alias, element_select, element_divide, placeholder, exp
if time_step > 0:
return past_value(x,
time_step=time_step,
initial_state=initial_state,
name=name)
elif time_step < 0:
return future_value(x,
time_step=-time_step,
initial_state=initial_state,
name=name)
else:
if name:
return alias(x, name)
else:
return x
def reduce_max(x, name=''):
"""
Get the max value in the sequence values
Args:
x: input sequence
name: the name of the operator
Returns:
The max value in the input sequence
"""
from ...ops import element_select, placeholder, greater
m = placeholder(shape=(1, ), dynamic_axes=x.dynamic_axes, name='max')
o = element_select(greater(x, future_value(m)), x, future_value(m))
rlt = o.replace_placeholders({m: sanitize_input(o)})
max_out = first(rlt)
return sanitize_input(max_out)
def future_value(x, initial_state=None, time_step=1, name=''):
'''
This function returns the future value w.r.t. ``x``. It is most often used when
creating RNNs. The resulting tensor has the same shape as the input but is
the next logical sample. The ``time_step`` parameter is the number of steps
to look into the future and is 1 by default. If there is no future value (i.e.
the current sample is the last one in the tensor) then the ``initial_state``
value is returned.
The initial state can be a constant (scalar or tensor), a learnable tensor
or input data (which has a batch dimension, as needed for sequence-to-sequence models).
Example:
>>> x = C.sequence.input_variable(shape=(3,2))
>>> # Create one sequence with 4 tensors of shape (3, 2)
>>> x0 = np.reshape(np.arange(24,dtype=np.float32),(1,4,3,2))
>>> y = C.sequence.future_value(x) # using initial state of 0 by default
>>> y.eval({x:x0})
[array([[[ 6., 7.],
[ 8., 9.],
[ 10., 11.]],
<BLANKLINE>
[[ 12., 13.],
[ 14., 15.],
[ 16., 17.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 0., 0.],
[ 0., 0.],
[ 0., 0.]]], dtype=float32)]
Args:
x: the tensor (or its name) from which the future value is obtained.
initial_state: tensor or scalar representing the initial value to be used when the input tensor is shifted in time.
time_step (int): the number of time steps to look into the future (default 1)
name (str, optional): the name of the Function instance in the network
Returns:
:class:`~cntk.ops.functions.Function`
'''
from cntk.internal import sanitize_dtype_cntk
from ...cntk_py import Constant
from cntk.cntk_py import future_value
if initial_state is None:
initial_state = Constant.scalar(sanitize_dtype_cntk(x.dtype), 0.0)
else:
initial_state = sanitize_input(initial_state)
x = sanitize_input(x)
return future_value(x, initial_state, time_step, name)
def future_value(x, initial_state=None, time_step=1, name=''):
'''
This function returns the future value w.r.t. ``x``. It is most often used when
creating RNNs. The resulting tensor has the same shape as the input but is
the next logical sample. The ``time_step`` parameter is the number of steps
to look into the future and is 1 by default. If there is no future value (i.e.
the current sample is the last one in the tensor) then the ``initial_state``
value is returned.
The initial state can be a constant (scalar or tensor), a learnable tensor
or input data (which has a batch dimension, as needed for sequence-to-sequence models).
Example:
>>> x = C.sequence.input_variable(shape=(3,2))
>>> # Create one sequence with 4 tensors of shape (3, 2)
>>> x0 = np.reshape(np.arange(24,dtype=np.float32),(1,4,3,2))
>>> y = C.sequence.future_value(x) # using initial state of 0 by default
>>> y.eval({x:x0})
[array([[[ 6., 7.],
[ 8., 9.],
[ 10., 11.]],
<BLANKLINE>
[[ 12., 13.],
[ 14., 15.],
[ 16., 17.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 0., 0.],
[ 0., 0.],
[ 0., 0.]]], dtype=float32)]
Args:
x: the tensor (or its name) from which the future value is obtained.
initial_state: tensor or scalar representing the initial value to be used when the input tensor is shifted in time.
time_step (int): the number of time steps to look into the future (default 1)
name (str, optional): the name of the Function instance in the network
Returns:
:class:`~cntk.ops.functions.Function`
'''
from cntk.internal import sanitize_dtype_cntk
from ...cntk_py import Constant
from cntk.cntk_py import future_value
if initial_state is None:
initial_state = Constant.scalar(sanitize_dtype_cntk(x.dtype), 0.0)
else:
initial_state = sanitize_input(initial_state)
x = sanitize_input(x)
return future_value(x, initial_state, time_step, name)
def delay(x, initial_state=None, time_step=1, name=''):
'''
This function combines :func:`~cntk.ops.sequence.past_value` and :func:`~cntk.ops.sequence.future_value` into a single function.
This is useful when the time_step is computed and can be positive, negative, or 0.
Args:
x: the tensor (or its name) from which the past value is obtained
initial_state: tensor or scalar representing the initial value to be used when the input tensor is shifted in time.
time_step (int): the number of time steps to look into the past, where negative values mean to look into the future, and 0 means a no-op (default 1).
name (str, optional): the name of the Function instance in the network
'''
from ...ops import alias
if time_step > 0:
return past_value (x, time_step= time_step, initial_state=initial_state, name=name)
elif time_step < 0:
return future_value(x, time_step=-time_step, initial_state=initial_state, name=name)
else:
if name:
return alias(x, name)
else:
return x
