Lucid3.0
module

init

15 members
lucid.nn.init

nn.init: parameter initialization functions. All functions operate in-place and return the tensor.

Functions

fnuniform_→ Tensor

Initialise `tensor` in-place with samples from a uniform distribution.

fnnormal_→ Tensor

Initialise `tensor` in-place with samples from a Gaussian distribution.

fnconstant_→ Tensor

Fill `tensor` in-place with a single scalar value.

fnones_→ Tensor

Fill `tensor` in-place with ones.

fnzeros_→ Tensor

Fill `tensor` in-place with zeros.

fneye_→ Tensor

Fill a 2-D `tensor` in-place with the identity matrix.

fnxavier_uniform_→ Tensor

Initialise `tensor` in-place with Xavier (Glorot) uniform initialisation.

fnxavier_normal_→ Tensor

Initialise `tensor` in-place with Xavier (Glorot) normal initialisation.

fnkaiming_uniform_→ Tensor

Initialise `tensor` in-place with Kaiming (He) uniform initialisation.

fnkaiming_normal_→ Tensor

Initialise `tensor` in-place with Kaiming (He) normal initialisation.

fntrunc_normal_→ Tensor

Initialise `tensor` in-place with truncated normal samples.

fnorthogonal_→ Tensor

Initialise `tensor` in-place with a (semi-)orthogonal matrix.

fnsparse_→ Tensor

Initialise a 2-D `tensor` in-place with a sparse random matrix.

fndirac_→ Tensor

Initialise a 3/4/5-D convolution weight in-place as a Dirac delta.

fncalculate_gain→ float

Return the recommended variance-preserving gain for an activation.