class

Log2Backward

Log2Backward

source edit

Python wrappers

pylucid.ops.log2function

Autograd node for the element-wise base-2 logarithm $y = \log_2(x)$ .

Saves the input x; the backward divides by $x \ln 2$ . The constant $\ln 2$ is materialised in grad_formula as a high-precision literal so accuracy is identical to scaling by 1 / log(2.0).

Math

\frac{\partial y}{\partial x} = \frac{1}{x \ln 2}, \qquad \frac{\partial \mathcal{L}}{\partial x} = \frac{1}{x \ln 2}\,\frac{\partial \mathcal{L}}{\partial y}, \quad x > 0.$$

Attributes

schema_v1OpSchema

Registered as "log2" with AmpPolicy::ForceFP32.

Notes

Dispatch: Accelerate vvlog2f/vvlog2 (CPU) / MLX log2 (GPU).

Static methods

static

dispatch

int dispatch(backend::IBackend & be, const int & a, const int & s, Dtype dt)

source edit

Forward — calls IBackend::log2 to compute $y = \log_2 x$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Backward — $\partial y/\partial x = 1/(x \ln 2)$ , scaled by grad_out.

class

Log2Backward

Log2Backward

source edit

Python wrappers

pylucid.ops.log2function

Autograd node for the element-wise base-2 logarithm $y = \log_2(x)$ .

Saves the input x; the backward divides by $x \ln 2$ . The constant $\ln 2$ is materialised in grad_formula as a high-precision literal so accuracy is identical to scaling by 1 / log(2.0).

Math

\frac{\partial y}{\partial x} = \frac{1}{x \ln 2}, \qquad \frac{\partial \mathcal{L}}{\partial x} = \frac{1}{x \ln 2}\,\frac{\partial \mathcal{L}}{\partial y}, \quad x > 0.$$

Attributes

schema_v1OpSchema

Registered as "log2" with AmpPolicy::ForceFP32.

Notes

Dispatch: Accelerate vvlog2f/vvlog2 (CPU) / MLX log2 (GPU).

Static methods

static

dispatch

int dispatch(backend::IBackend & be, const int & a, const int & s, Dtype dt)

source edit

Forward — calls IBackend::log2 to compute $y = \log_2 x$ .

Methods

grad_formula

int grad_formula(const int & g)

source edit

Backward — $\partial y/\partial x = 1/(x \ln 2)$ , scaled by grad_out.