def softmax(x, name = ''): """ Compute softmax along with a squence values """ from ...ops import element_divide, exp x_exp = exp((x-broadcast_as(reduce_max(x), x))*10) x_softmax = element_divide(x_exp, broadcast_as(reduce_sum(x_exp), x), name = name) return x_softmax
def broadcast_as(operand, broadcast_as_operand, name=''):
'''
Creates a sequence out of a non-sequence by endowing the `operand`
with dynamic axes of the same type as the `broadcast_as_operand`
and broadcasting the value of the `operand` along those dynamic axes.
Example:
>>> import cntk.ops as C
>>> import numpy as np
>>> x = C.input_variable(shape=(3,2))
>>> t = C.sequence.last(x)
>>> b = C.sequence.is_first(x)
>>> y = C.sequence.broadcast_as(t, b)
>>> x0 = np.reshape(np.arange(24.0,dtype=np.float32),(4,3,2))
>>> y.eval({x:x0})
array([[[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]]]], dtype=float32)
Args:
operand: the symbolic tensor whose value will be broadcast
broadcast_as_operand: the symbolic tensor whose dynamic axes will
be used to broadcast the operand
name (str): the name of the node in the network
Returns:
:class:`cntk.Function`
'''
from cntk.cntk_py import broadcast_as
operand = sanitize_input(operand, get_data_type(operand))
broadcast_as_operand = sanitize_input(
broadcast_as_operand, get_data_type(broadcast_as_operand))
return broadcast_as(operand, broadcast_as_operand, name)
def broadcast_as(operand, broadcast_as_operand, name=''):
'''
Creates a sequence out of a non-sequence by endowing the ``operand``
with dynamic axes of the same type as the ``broadcast_as_operand``
and broadcasting the value of the ``operand`` along those dynamic axes.
Example:
>>> import cntk.ops as C
>>> import numpy as np
>>> x = C.input_variable(shape=(3,2))
>>> t = C.sequence.last(x)
>>> b = C.sequence.is_first(x)
>>> y = C.sequence.broadcast_as(t, b)
>>> x0 = np.reshape(np.arange(24.0,dtype=np.float32),(4,3,2))
>>> y.eval({x:x0})
array([[[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]],
<BLANKLINE>
[[ 18., 19.],
[ 20., 21.],
[ 22., 23.]]]], dtype=float32)
Args:
operand: the symbolic tensor whose value will be broadcast
broadcast_as_operand: the symbolic tensor whose dynamic axes will
be used to broadcast the operand
name (str): the name of the node in the network
Returns:
:class:`cntk.Function`
'''
from cntk.cntk_py import broadcast_as
operand = sanitize_input(operand, get_data_type(operand))
broadcast_as_operand = sanitize_input(
broadcast_as_operand, get_data_type(broadcast_as_operand))
return broadcast_as(operand, broadcast_as_operand, name)
def broadcast_as(operand, broadcast_as_operand, name=''):
'''
TBA
Example:
TBA
Args:
operand: the symbolic tensor operand denoting a tensor
broadcast_as_operand: the symbolic tensor operand denoting a sequence per whose layout the main operand id to be broadcast
name (str): the name of the node in the network
Returns:
:class:`cntk.Function`
'''
from cntk.cntk_py import broadcast_as
operand = sanitize_input(operand, get_data_type(operand))
broadcast_as_operand = sanitize_input(
broadcast_as_operand, get_data_type(broadcast_as_operand))
return broadcast_as(operand, broadcast_as_operand, name)
def broadcast_as(operand, broadcast_as_operand, name=''):
'''
TBA
Example:
TBA
Args:
operand: the symbolic tensor operand denoting a tensor
broadcast_as_operand: the symbolic tensor operand denoting a sequence per whose layout the main operand id to be broadcast
name (str): the name of the node in the network
Returns:
:class:`cntk.Function`
'''
from cntk.cntk_py import broadcast_as
operand = sanitize_input(operand, get_data_type(operand))
broadcast_as_operand = sanitize_input(broadcast_as_operand,
get_data_type(broadcast_as_operand))
return broadcast_as(operand, broadcast_as_operand, name)
