sum_op

→TensorImplPtr

int sum_op(const int & a, const int & axes, bool keepdims)

source edit

Python wrappers

pylucid.ops.sumfunction

Reduce a by addition along axes.

Promotes bool / int8 / int16 / int32 inputs to int64 before dispatch so the sum has overflow headroom matching the reference framework. Routes the forward through SumBackward which also installs the autograd edge when a->requires_grad() is true.

See Also

SumBackward — backward node.

Parameters

aTensorImplPtr

Input tensor of any shape and any real dtype.

axesstd::vector<int>

Axes to reduce. May be empty to reduce all axes. Negative indices wrap around ndim.

keepdimsbool

If true, the reduced dimensions are kept as size-1 entries; otherwise they are collapsed.

Returns

TensorImplPtr

Output tensor. Shape is the input shape with the reduced axes either removed (default) or set to $1$ (keepdims=true).