pow_op

→TensorImplPtr

int pow_op(const int & a, const int & b)

Compute the elementwise tensor-tensor power $y = a^b$ with broadcasting and autograd support.

aTensorImplPtr

Base tensor.

bTensorImplPtr

Exponent tensor. Broadcast-compatible with a.

TensorImplPtr

Tensor of shape broadcast(a.shape, b.shape) holding $a^b$ .

Always computed in FP32 (see PowBackward schema). When grad-tracking is on, registers a PowBackward node which saves both inputs.

>>> auto y = pow_op(a, b);  // y[i] = a[i] ** b[i] with broadcasting

Loading…