def _CoshGrad(op, grad):
"""Returns grad * sinh(x)."""
x = op.inputs[0]
with ops.control_dependencies([grad]):
x = math_ops.conj(x)
return grad * math_ops.sinh(x)
def _AcoshGrad(op, grad):
"""Returns grad * 1/sinh(y)."""
y = op.outputs[0]
with ops.control_dependencies([grad]):
y = math_ops.conj(y)
return grad / math_ops.sinh(y)
def _forward(self, x): return math_ops.sinh((math_ops.asinh(x) + self.skewness) * self.tailweight)
def _inverse(self, y): return math_ops.sinh(math_ops.asinh(y) / self.tailweight - self.skewness)
def _AcoshGrad(op, grad):
"""Returns grad * 1/sinh(y)."""
y = op.outputs[0]
with ops.control_dependencies([grad]):
y = math_ops.conj(y)
return grad / math_ops.sinh(y)
def _CoshGrad(op, grad):
"""Returns grad * sinh(x)."""
x = op.inputs[0]
with ops.control_dependencies([grad]):
x = math_ops.conj(x)
return grad * math_ops.sinh(x)
def _inverse(self, y): return math_ops.sinh(math_ops.asinh(y) / self.tailweight - self.skewness)
def _forward(self, x): return math_ops.sinh((math_ops.asinh(x) + self.skewness) * self.tailweight)
def exp_map(self, v, x):
# eq (9), v is the gradient
vnorm = math_ops.sqrt(self.lorentz_scalar_product(v, v))
return math_ops.cosh(vnorm) * x + math_ops.sinh(vnorm) * v / vnorm
