def visit(self, op, x):
"""
TODO.
Arguments:
op: TODO
Returns:
TODO
"""
if x.is_scalar and x.is_constant:
self.replace_op(op, constant(-x.const))
def visit(self, op, x):
"""
TODO.
Arguments:
op: TODO
Returns:
TODO
"""
if x.is_scalar and x.is_constant:
val = power(x.const, op.reduction_axes.size)
self.replace_op(op, constant(val))
def visit(self, op):
"""
TODO.
Arguments:
op: TODO
Returns:
TODO
"""
x, = op.args
if x.is_scalar and x.is_constant:
val = x.const * op.reduction_axes.size
self.replace_op(op, constant(val))
def visit(self, op, x):
"""
If x is filled with the same value, then replace the prod op
with `power`.
Arguments:
op: TODO
Returns:
TODO
"""
if x.is_scalar and x.is_constant:
val = power(x.const, op.reduction_axes.size)
self.replace_op(op, constant(val))
def visit(self, op, x):
"""
If x is filled with the same value, then replace the prod op
with x.const ** reduction_axes.size.
Arguments:
op: TODO
Returns:
TODO
"""
if x.is_scalar and x.is_constant:
val = x.const**op.reduction_axes.size
self.replace_op(op, constant(val, op.axes))
def _deriv(dependent, independent, error=constant(1)):
"""
Computes the operation for [dDependent/dIndependent](error=1).
The derivative is a multi-linear function.
Args:
dependent (TensorOp): Dependent op.
independent(TensorOp): Independent op.
error (TensorOp, optional): The tensor holding the error where the
derivative will be computed at. Must have the same axes as dependent.
Returns:
TensorOp: Derivative applied to error. Has axes of independent.
"""
if not error.axes.has_same_axes(dependent.axes):
raise ValueError(
"Dependent and error must have the same set of axes")
adjoints = dependent.forwarded.adjoints(error)
if independent.forwarded.tensor not in adjoints:
return constant(0, independent.axes)
adjoint = adjoints[independent.forwarded.tensor]
return broadcast(adjoint.forwarded, axes=independent.axes)