Gradien

Appearance

Layers Module ​

Layers are the building blocks of neural networks. They store learnable parameters (weights and biases).

NN.Linear ​

Applies a linear transformation to the incoming data: y = x * W^T + b.

lua
(
    inFeatures: number,
    outFeatures: number,
    initializer: ((fanIn, fanOut) -> number)?
) -> Module
lua
local layer = Gradien.NN.Linear(128, 64)

NN.Sequential ​

A container that chains modules together. Data flows through them in the order they are defined.

lua
(layers: { Module | (Tensor)->Tensor }) -> Module
lua
local model = Gradien.NN.Sequential({
    Gradien.NN.Linear(10, 20),
    Gradien.NN.Linear(20, 5)
})

NN.Conv2d ​

Applies a 2D convolution over an input signal composed of several input planes.

lua
(C_in: number, C_out: number, KH: number, KW: number) -> Module
lua
local conv = Gradien.NN.Conv2d(3, 64, 3, 3) -- 3 input channels, 64 output, 3x3 kernel

NN.MaxPool2d ​

Applies a 2D max pooling over an input signal.

lua
(KH: number, KW: number, stride: number) -> Module
lua
local pool = Gradien.NN.MaxPool2d(2, 2, 2) -- 2x2 kernel, stride 2

NN.AvgPool2d ​

Applies a 2D average pooling over an input signal.

lua
(KH: number, KW: number, stride: number) -> Module
lua
local pool = Gradien.NN.AvgPool2d(2, 2, 2) -- 2x2 kernel, stride 2

NN.ConvTranspose2d ​

Applies a 2D transposed convolution operator over an input image composed of several input planes.

lua
(C_in: number, C_out: number, KH: number, KW: number) -> Module
lua
local convT = Gradien.NN.ConvTranspose2d(64, 3, 3, 3) -- 64 input channels, 3 output, 3x3 kernel

NN.GroupNorm ​

Applies Group Normalization over a mini-batch of inputs. Divides channels into groups and normalizes within each group independently.

lua
(num_groups: number, num_channels: number, eps: number?) -> Module
lua
local gn = Gradien.NN.GroupNorm(8, 64) -- 8 groups, 64 channels

Parameters:

  • num_groups (number): Number of groups to divide channels into. Must divide num_channels evenly.
  • num_channels (number): Number of channels expected in the input.
  • eps (number, optional): Small value added to variance for numerical stability. Default: 1e-5.