Math & logic

Elementwise numeric, trigonometric, activation, and boolean operations.

Arithmetic

Elementwise numeric operations: rounding, powers, logs, exponentials, affine transforms, sign, and special functions.

Function	Description
Absolute value	Absolute value of each element.
Add	Elementwise sum of two aligned streams (x + y).
Ceiling	Round each element toward positive infinity.
Clip to bounds	Bound each element below and/or above.
Cube	x cubed (faster than Power(3)).
Div	Elementwise quotient of two aligned streams (x / y).
Error function	Gauss error function.
Complementary error function	Complementary error function (1 - erf).
Exponential	e to the power of each element.
Floor	Round each element toward negative infinity.
Identity	Pass-through (y = x).
Linear (affine)	Affine transform: scale * x + shift.
Linear (two-input affine)	Two-input affine combination: a*x + b*y + c.
Natural logarithm	Natural logarithm of each element.
Mul	Elementwise product of two aligned streams (x * y).
Negative Part	Negative part of x: max(-x, 0).
Positive Part	Positive part of x: max(x, 0).
Power	x raised to a fixed exponent p.
Round (banker's rounding)	Round each element to the nearest integer (half-to-even).
Sign	Sign of each element: -1, 0, or +1.
Square root	Square root of each element.
Square	x squared (faster than Power(2)).
Sub	Elementwise difference of two aligned streams (x - y).

Trigonometry & geometry

Angles, inverse trigonometric functions, coordinate conversions, and distances.

Function	Description
Arccosine	Inverse cosine of each element (radians, input in [-1, 1]).
Arcsine	Inverse sine of each element (radians, input in [-1, 1]).
Arctangent	Inverse tangent of each element (radians).
Two-argument arctangent	Signed angle of (x, y) from the positive x-axis (numpy.arctan2 order).
Cartesian to polar	Convert (x, y) to (r, theta).
Cosine	Cosine of each element (radians).
Hypotenuse	Euclidean distance sqrt(x^2 + y^2), numerically stable.
Polar to Cartesian	Convert (r, theta) to (x, y).
Sine	Sine of each element (radians).
Tanh	Hyperbolic tangent.

Activations

Neural-network activation functions.

Function	Description
ELU	Exponential linear unit.
ReLU	Rectified linear unit: max(0, x).
SELU	Scaled exponential linear unit (self-normalizing networks).
Sigmoid	Logistic sigmoid: 1 / (1 + exp(-x)).
Softsign	Softsign: x / (1 + |x|).
Tanh	Hyperbolic tangent.

Logic & comparison

Boolean masks, comparisons, and conditional selection. Outputs are 1.0 (true) or 0.0 (false).

Function	Description
And	Returns 1.0 if both inputs are nonzero, else 0.0. NaN in either input yields NaN.
Equal	Returns 1.0 if a == b, else 0.0. NaN in either input yields NaN.
GreaterEqual	Returns 1.0 if a >= b, else 0.0. NaN in either input yields NaN.
GreaterThan	Returns 1.0 if a > b, else 0.0. NaN in either input yields NaN.
IsFinite	Returns 1.0 for finite values, 0.0 for NaN or inf. Does not propagate NaN.
IsNan	Returns 1.0 if the input is NaN, else 0.0. Does not propagate NaN.
LessEqual	Returns 1.0 if a <= b, else 0.0. NaN in either input yields NaN.
LessThan	Returns 1.0 if a < b, else 0.0. NaN in either input yields NaN.
Not	Returns 0.0 if input is nonzero, 1.0 if zero. NaN propagates.
NotEqual	Returns 1.0 if a != b, else 0.0. NaN in either input yields NaN.
Or	Returns 1.0 if either input is nonzero, else 0.0. NaN in either input yields NaN.
Where	Returns a if mask is nonzero, b otherwise. NaN mask yields NaN.