def forward(self, inputs):
xp = cuda.get_array_module(*inputs)
x, t = inputs
if xp is np:
self.y, = softmax.Softmax().forward((x, ))
loss = softmax_crossentropy_cpu_funcs.crossentropy(self.y,
t) / t.shape[0]
else:
self.y, = softmax.Softmax(self.use_cudnn).forward((x, ))
loss = _get_crossentropyloss_gpu(self.y, t)
return loss.reshape(()),
def forward_gpu(self, inputs):
x, t = inputs
self.y, = softmax.Softmax(self.use_cudnn).forward_gpu((x,))
ret = cuda.reduce(
'int* t, float* y, int n_channel', '-log(y[i * n_channel + t[i]])',
'a+b', '0', 'crossent_fwd', numpy.float32
)(t, self.y, self.y.shape[1])
ret /= t.size
return ret,
def forward_gpu(self, inputs):
cupy = cuda.cupy
x, t = inputs
self.y, = softmax.Softmax(self.use_cudnn).forward((x, ))
if getattr(self, 'normalize', True):
count = x.size // x.shape[1]
else:
count = x.shape[0]
y = cupy.rollaxis(self.y, 1, self.y.ndim)
ret = cuda.reduce('S t, raw T y, int32 n_channel, T inv_count',
'T out', 'log(y[_j * n_channel + t])', 'a + b',
'out = a * inv_count', '0',
'crossent_fwd')(t, y.reduced_view(), y.shape[-1],
-1.0 / count)
return ret,
def forward_cpu(self, inputs):
x, t = inputs
self.y, = softmax.Softmax().forward((x, ))
yd = numpy.rollaxis(self.y, 1)
yd = yd.reshape(len(yd), -1).T
p = yd[six.moves.range(t.size), t.flat]
# deal with the case where the SoftmaxCrossEntropy is
# unpickled from the old version
if getattr(self, 'normalize', True):
count = x.size // x.shape[1]
else:
count = x.shape[0]
y = numpy.log(p).sum(keepdims=True) * (-1.0 / count)
return y.reshape(()),
def forward_gpu(self, inputs):
cupy = cuda.cupy
x, t = inputs
self.y, = softmax.Softmax(self.use_cudnn).forward((x, ))
n_unit = int(numpy.prod(self.y.shape[2:]))
if getattr(self, 'normalize', True):
count = t.shape[0] * n_unit
else:
count = t.shape[0]
y = cupy.rollaxis(self.y, 1, len(self.y))
ret = cuda.reduce('S t, raw T y, int32 n_channel, T inv_count',
'T out', 'log(y[_j * n_channel + t])', 'a + b',
'out = a * inv_count', '0',
'crossent_fwd')(t, y, y.shape[-1], -1.0 / count)
return ret,
def forward_gpu(self, inputs):
x, t = inputs
self.y, = softmax.Softmax(self.use_cudnn).forward_gpu((x,))
n_unit = int(numpy.prod(self.y.shape[2:]))
# the map_expr is equivalent to the pseudo code -log(y[n, c, m]),
# where n = i / n_unit, c = t[i], and m = i % n_unit
ret = cuda.reduce(
'int* t, float* y, int n_channel, int n_unit',
'-log(y[n_unit * ((i / n_unit) * n_channel + t[i])'
' + (i % n_unit)])',
'a+b', '0', 'crossent_fwd', numpy.float32
)(t, self.y, self.y.shape[1], n_unit)
if getattr(self, 'normalize', True):
n_unit = int(numpy.prod(self.y.shape[2:]))
count = t.shape[0] * n_unit
else:
count = t.shape[0]
ret /= count
return ret,
def forward_gpu(self, inputs):
cupy = cuda.cupy
x, t = inputs
self.y, = softmax.Softmax(self.use_cudnn).forward((x,))
if getattr(self, 'normalize', True):
count = float((t != self.ignore_label).sum())
else:
count = t.shape[0]
self.count = count
if count == 0:
return cupy.zeros((), dtype=x.dtype),
y = cupy.rollaxis(self.y, 1, self.y.ndim)
ret = cuda.reduce(
'S t, raw T y, int32 n_channel, T inv_count', 'T out',
't == -1 ? 0 : log(y[_j * n_channel + t])',
'a + b', 'out = a * inv_count', '0', 'crossent_fwd'
)(t, y.reduced_view(), y.shape[-1], -1.0 / count)
return ret,
def forward_cpu(self, inputs):
x, t = inputs
self.y, = softmax.Softmax().forward((x,))
yd = numpy.rollaxis(self.y, 1)
yd = yd.reshape(len(yd), -1).T
p = yd[six.moves.range(t.size), numpy.maximum(t.flat, 0)]
# deal with the case where the SoftmaxCrossEntropy is
# unpickled from the old version
if getattr(self, 'normalize', True):
count = (t != self.ignore_label).sum()
else:
count = x.shape[0]
self.count = count
if count == 0:
return numpy.zeros((), dtype=x.dtype),
y = (numpy.log(p) * (t.flat != self.ignore_label)).sum(keepdims=True) \
* (-1.0 / count)
return y.reshape(()),
def forward_cpu(self, inputs):
x, t = inputs
self.y, = softmax.Softmax().forward_cpu((x,))
p = self.y[six.moves.range(len(t)), t]
y = -numpy.log(p).sum(keepdims=True) / t.size
return y.reshape(()),
