from chainer import cuda
from chainer import function
from chainer.utils import type_check
class LogSumExp(function.Function):
def __init__(self, axis=None):
if axis is None:
self.axis = None
elif isinstance(axis, int):
self.axis = (axis,)
elif isinstance(axis, tuple) and all(isinstance(a, int) for a in axis):
if len(set(axis)) != len(axis):
raise ValueError('duplicate value in axis: ({})'.format(
', '.join(map(str, axis))))
self.axis = axis
else:
raise TypeError('None, int or tuple of int are required')
def check_type_forward(self, in_types):
type_check.expect(
in_types.size() == 1,
in_types[0].dtype.kind == 'f',
)
if self.axis is not None:
for axis in self.axis:
if axis >= 0:
type_check.expect(
axis < in_types[0].ndim,
)
else:
type_check.expect(
-axis - 1 < in_types[0].ndim,
)
def forward(self, inputs):
xp = cuda.get_array_module(*inputs)
x, = inputs
m = x.max(axis=self.axis, keepdims=True)
y = x - m
xp.exp(y, out=y)
y_sum = y.sum(axis=self.axis)
self.y = xp.asarray(xp.log(y_sum) + m.reshape(y_sum.shape))
return self.y,
def backward(self, inputs, grads):
xp = cuda.get_array_module(*inputs)
x, = inputs
gy, = grads
y = self.y
if self.axis is not None:
actual_axis = []
for axis in self.axis:
if axis < 0:
axis = len(x.shape) + axis
actual_axis.append(axis)
for axis in sorted(actual_axis):
gy = xp.expand_dims(gy, axis=axis)
y = xp.expand_dims(y, axis=axis)
gx = gy * xp.exp(x - y)
return gx,
[docs]def logsumexp(x, axis=None):
"""Log-sum-exp of array elements over a given axis.
This function calculates logarithm of sum of exponential of array elements.
.. math::
y_i = \\log\\left(\\sum_j \\exp(x_{ij})\\right)
Args:
x (~chainer.Variable): Elements to log-sum-exp.
axis (None, int, or tuple of int): Axis which a sum is performed.
The default (axis = None) is perform a sum over all the dimensions
of the input array.
Returns:
~chainer.Variable: Output variable.
"""
return LogSumExp(axis)(x)