Struct Linear 

pub struct Linear { /* private fields */ }

A linear transformation layer (fully connected layer).

The Linear layer applies a linear transformation to the incoming data: y = xW^T + b. This is one of the most fundamental building blocks in neural networks.

Parameters

• weight: Learnable weight matrix of shape [out_features, in_features]
• bias: Optional learnable bias vector of shape [out_features]

Input/Output Shapes

• Input: [..., in_features] - Can be 2D or 3D
• Output: [..., out_features] - Same number of dimensions as input

Example

use trustformers_core::layers::Linear;
use trustformers_core::tensor::Tensor;
use trustformers_core::traits::Layer;

// Create a linear layer: 768 → 3072
let linear = Linear::new(768, 3072, true);

// Apply to 2D input: [seq_len, in_features]
let input_2d = Tensor::randn(&[128, 768])?;
let output_2d = linear.forward(input_2d)?;  // Shape: [128, 3072]

// Apply to 3D input: [batch, seq_len, in_features]
let input_3d = Tensor::randn(&[4, 128, 768])?;
let output_3d = linear.forward(input_3d)?;  // Shape: [4, 128, 3072]

Implementations

impl Linear

pub fn new(in_features: usize, out_features: usize, bias: bool) -> Self

Creates a new linear layer.

Arguments

• in_features - Size of each input sample
• out_features - Size of each output sample
• bias - Whether to include a learnable bias

Returns

A new Linear layer with randomly initialized weights using a normal distribution, and bias initialized to zeros if enabled.

Example

use trustformers_core::layers::Linear;

// Linear layer without bias
let linear1 = Linear::new(512, 1024, false);

// Linear layer with bias
let linear2 = Linear::new(512, 1024, true);

pub fn new_with_device( in_features: usize, out_features: usize, bias: bool, device: Device, ) -> Self

Creates a new linear layer with specified device.

Arguments

• in_features - Size of each input sample
• out_features - Size of each output sample
• bias - Whether to include a learnable bias
• device - Device to use for computations (CPU, Metal, CUDA, etc.)

Returns

A new Linear layer with randomly initialized weights.

Example

use trustformers_core::layers::Linear;
use trustformers_core::Device;

// Create a linear layer on Metal GPU
let linear = Linear::new_with_device(768, 3072, true, Device::metal_if_available(0));

pub fn device(&self) -> Device

Returns the device this layer uses for computations.

pub fn to_device(self, device: Device) -> Self

Move this layer to a different device.

Arguments

• device - Target device

Returns

Self with updated device.

pub fn set_weight(&mut self, weight: Tensor) -> Result<()>

Sets the weight matrix for this layer.

Arguments

• weight - The new weight tensor, must have shape [out_features, in_features]

Returns

Ok(()) if successful.

Note

This method is typically used when loading pretrained weights.

pub fn set_bias(&mut self, bias: Tensor) -> Result<()>

Sets the bias vector for this layer.

Arguments

• bias - The new bias tensor, must have shape [out_features]

Returns

Ok(()) if successful.

Note

This will enable bias even if the layer was created without bias.

pub fn weight(&self) -> &Tensor

Returns a reference to the weight matrix.

Returns

A reference to the weight tensor of shape [out_features, in_features].

pub fn bias(&self) -> Option<&Tensor>

Returns a reference to the bias vector if present.

Returns

Some(&bias) if bias is enabled, None otherwise.

pub fn parameter_count(&self) -> usize

Returns the total number of learnable parameters in this layer.

Returns

The total parameter count including weights and bias (if present).

Trait Implementations

impl Clone for Linear

fn clone(&self) -> Linear

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Linear

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Layer for Linear

type Input = Tensor

type Output = Tensor

fn forward(&self, input: Self::Input) -> Result<Self::Output>

Performs the forward computation of this layer.