def forward_cpu(self, x: tensor) -> tensor:
data = np.tanh(x.host_data)
if self.inplace:
self.cache = [x, x]
x.host_data = data
return x
else:
y = zeros_like(x)
self.cache = [x, y]
y.host_data = data
return y
def forward_cpu(self, x: tensor) -> tensor:
tmp = x.host_data > 0
data = x.host_data * tmp
if self.inplace:
self.cache = [x, tmp, x]
x.host_data = data
return x
else:
y = zeros_like(x)
self.cache = [x, tmp, y]
y.host_data = data
return y
def _backward_cpu(self, dx: tensor, dy: tensor) -> tensor:
dx.host_data = np.transpose(dy.host_data, self.axes)
return dx
def softmax_util_cpu(x: tensor, y: tensor) -> tensor:
eX = np.exp((x.host_data.T - np.max(x.host_data, axis=1)).T)
y.host_data = (eX.T / eX.sum(axis=1)).T
return y
def _backward_cpu(self, dx: tensor, dy: tensor) -> tensor:
dx.host_data = dy.host_data / (1 - self.prob)
dx[self.mask.host_data] = 0
return dx