def past_value(shape, x, time_step=1, default_hidden_activation=0.1, name=None):
"""
This function returns the past value wrt `x`. It is most often used when
creating RNNs. The resulting tensor has the same shape as the input but is
the previous logical sample. The `time_step` parameter is the number of steps
to look into the past and is 1 by default. If there is no past value (i.e.
the current sample is the first one in the tensor) then the `default_hidden_activation`
value is returned which is 0.1 by default.
Example:
>>> data = np.array([[1,2,3,4],[5,6,7,8],[9,10,11,12]])
>>> t = C.dynamic_axis(name='t')
>>> x = C.input_numpy([data], dynamic_axis=t)
>>> with C.LocalExecutionContext('past_value') as ctx:
... print(ctx.eval(C.past_value(0, x)))
[array([[ 0.1, 0.1, 0.1, 0.1],
[ 1. , 2. , 3. , 4. ],
[ 5. , 6. , 7. , 8. ]])]
Args:
shape: dimensions of the input `x`
x: the tensor from which the past value is obtained
time_step: the number of time steps to look into the past (default 1)
default_hidden_activation: the default value to use when no past value
is available (default 0.1)
Returns:
:class:`cntk.graph.ComputationNode`
"""
from cntk.ops.cntk1 import PastValue
op = PastValue(shape, x, time_step, default_hidden_activation, name = name)
wrap_numpy_arrays(op)
op.rank = get_rank(shape)
return op
def past_value(dims, x, time_step=1, default_hidden_activation=0.1, name=None):
"""
This function returns the past value wrt `x`. It is most often used when
creating RNNs. The resulting tensor has the same shape as the input but is
the previous logical sample. The `time_step` parameter is the number of steps
to look into the past and is 1 by default. If there is no past value (i.e.
the current sample is the first one in the tensor) then the `default_hidden_activation`
value is returned which is 0.1 by default.
Example:
>>> data = np.array([[1,2,3,4],[5,6,7,8],[9,10,11,12]])
>>> t = C.dynamic_axis(name='t')
>>> x = C.input_numpy([data], dynamic_axis=t)
>>> with C.LocalExecutionContext('past_value') as ctx:
... print(ctx.eval(C.past_value(0, x)))
[array([[ 0.1, 0.1, 0.1, 0.1],
[ 1. , 2. , 3. , 4. ],
[ 5. , 6. , 7. , 8. ]])]
Args:
dims: dimensions of the input `x`
x: the tensor from which the past value is obtained
time_step: the number of time steps to look into the past (default 1)
default_hidden_activation: the default value to use when no past value
is available (default 0.1)
Returns:
:class:`cntk.graph.ComputationNode`
"""
from cntk.ops.cntk1 import PastValue
return PastValue(dims, x, time_step, default_hidden_activation, name=name)
def test_loose_coupling():
from cntk.ops.cntk1 import PastValue
dh = PastValue(1, 'outnode')
out = Times(dh, Constant(2), var_name='outnode')
expected = ['v0 = PastValue(1, outnode, timeStep=1, defaultHiddenActivation=0.1)',
'v1 = Constant(2, rows=1, cols=1)',
'outnode = Times(v0, v1, outputRank=1)']
description, has_inputs, readers = out.to_config()
assert _to_list(description) == expected
def test_loose_coupling():
from cntk.ops.cntk1 import PastValue
dh = PastValue(1, 'outnode')
out = times(dh, constant(2), name='outnode')
expected = [
'v0 = PastValue(1, outnode, timeStep=1, defaultHiddenActivation=0.1)',
"v1 = ParameterTensor(1, learningRateMultiplier=0.0, init='fromLiteral', initValueScale=1, value=0, initFromFilePath='', initFromLiteral='2.0000",
"', initOnCPUOnly=true, randomSeed=-1)",
'outnode = Times(v0, v1, outputRank=1)'
]
description, has_inputs, readers = out.to_config()
assert _to_list(description) == expected