def backward(self, inputs, grad_outputs):
xp = cuda.get_array_module(*inputs)
x, t = inputs
gloss = grad_outputs[0]
y, = sigmoid.Sigmoid().forward((x,))
if self.reduce == 'mean':
gx = xp.divide(
gloss * self.ignore_mask * (y - t), self.count,
dtype=y.dtype)
else:
gx = (gloss * self.ignore_mask * (y - t)).astype(y.dtype)
return gx, None
def forward(self, inputs):
self.retain_inputs((0, 1))
xp = cuda.get_array_module(*inputs)
x, gy = inputs
y, = sigmoid.Sigmoid().forward((x, ))
if self.reduce == 'mean':
gx = xp.divide(gy * self.ignore_mask * (y - self.t),
self.count,
dtype=y.dtype)
else:
gx = (gy * self.ignore_mask * (y - self.t)).astype(y.dtype)
return gx, None
def forward(self, inputs):
self.retain_inputs((0, 1))
xp = backend.get_array_module(*inputs)
x, gy = inputs
y, = sigmoid.Sigmoid().forward((x, ))
if self.reduce == 'mean':
# TODO(takagi): Fix to perform division in a specific dtype. See
# cupy/cupy#1534.
gx = xp.divide(gy * self.ignore_mask * (y - self.t),
self.count).astype(y.dtype)
else:
gx = (gy * self.ignore_mask * (y - self.t)).astype(y.dtype)
return gx,