Developer Guide for Intel® Data Analytics Acceleration Library 2019 Update 1
A truncated Gaussian initializer is an initializer algorithm to initialize a p-dimensional tensor X ∈ R n1 x ... x np with variates that have the Gaussian probability density function with mean μ and standard deviation σ and belong to the truncation range [a, b], where b > a, μ,σ ∈ R, and σ > 0.
The bounds of the truncation range define the following cases:
No truncation: a = -∞, b = +∞
Left truncation: a > -∞, b = +∞
Right truncation: a = -∞, b < +∞
Two-sided truncation: a > -∞, b < +∞
In addition to common parameters of the initializer interface, a truncated Gaussian initializer has the following parameters:
Parameter |
Default Value |
Description |
|
---|---|---|---|
algorithmFPType |
float |
The floating-point type that the algorithm uses for intermediate computations. Can be float or double. |
|
method |
defaultDense |
Inverse cumulative distribution function (CDF) transform method. Performance-oriented computation method, the only method supported by the algorithm. |
|
mean |
0 |
The mean μ. |
|
sigma |
1 |
The standard deviation σ. |
|
a |
mean - 2*sigma |
The left bound of the truncation range. If it is set to the negative infinity, there is no left truncation. |
|
b |
mean + 2*sigma |
The right bound of the truncation range. If it is set to the positive infinity, there is no right truncation. |