def backward_impl(self, inputs, outputs, prop_down, accum):
# inputs: [inputs_fwd_graph] + [inputs_bwd_graph] or
# [inputs_fwd_graph] + [outputs_fwd_graph] + [inputs_bwd_graph]
# Inputs
x0 = inputs[0].data
dy = inputs[1].data
# Outputs
dx0 = outputs[0].data
# Grads of inputs
g_x0 = inputs[0].grad
g_dy = inputs[1].grad
# Grads of outputs
g_dx0 = outputs[0].grad
# Compute
val = self.forward_func.info.args["val"]
if prop_down[0] or prop_down[1]:
cv = F.constant(val, x0.shape)
if not nn.get_auto_forward():
cv.forward()
log_v = F.log(cv.data)
if prop_down[0]:
if accum[0]:
g_x0 += g_dx0 * dy * F.r_pow_scalar(x0, val) * log_v**2.0
else:
g_x0.copy_from(g_dx0 * dy * F.r_pow_scalar(x0, val) *
log_v**2.0)
if prop_down[1]:
if accum[1]:
g_dy += g_dx0 * F.r_pow_scalar(x0, val) * log_v
else:
g_dy.copy_from(g_dx0 * F.r_pow_scalar(x0, val) * log_v)
def __rpow__(self, other):
"""
Element-wise power function.
Part of the implementation of the power operator expression.
Args:
other (float or ~nnabla.Variable): Internally calling
:func:`~nnabla.functions.pow2` or
:func:`~nnabla.functions.r_pow_scalar` according to the
type.
Returns: :class:`nnabla.Variable`
"""
import nnabla.functions as F
if isinstance(other, Variable):
return F.pow2(other, self)
return F.r_pow_scalar(self, other)