Enum Layer 

pub enum Layer {
    Linear {
        name: String,
        in_features: usize,
        out_features: usize,
        weight: Vec<Vec<f32>>,
        bias: Vec<f32>,
    },
    Conv2d {
        name: String,
        in_channels: usize,
        out_channels: usize,
        kernel_size: usize,
        padding: usize,
        stride: usize,
        weight: Vec<Vec<Vec<Vec<f32>>>>,
        bias: Vec<f32>,
    },
    Activation {
        name: String,
        activation: Activation,
    },
    Flatten {
        name: String,
    },
    Reshape {
        name: String,
        shape: Vec<usize>,
    },
    Residual {
        name: String,
        layers: Vec<Layer>,
    },
}

Layer type definition for the universal inference system

Variants

Linear

Fully connected (linear) layer

Fields

name: String

Human-readable layer name, propagated from the exporter (e.g. "fc1"). Used only for diagnostics / pretty-printing.

in_features: usize

Input feature count. Must match the size of the incoming 1D tensor or Layer::forward returns an error.

out_features: usize

Output feature count, equal to weight.len() and bias.len().

weight: Vec<Vec<f32>>

Dense weight matrix, shape [out_features, in_features].

bias: Vec<f32>

Per-output bias, shape [out_features].

Conv2d

2D Convolutional layer

Fields

name: String

Human-readable layer name (e.g. "conv1").

in_channels: usize

Expected channel depth of the input tensor.

out_channels: usize

Channel depth of the produced tensor (= weight.len()).

kernel_size: usize

Side length of the square kernel; both height and width.

padding: usize

Zero-padding applied symmetrically on all four sides before the kernel sweep.

stride: usize

Step between successive kernel positions along both axes.

weight: Vec<Vec<Vec<Vec<f32>>>>

Kernel weights, shape [out_channels, in_channels, kernel_size, kernel_size].

bias: Vec<f32>

Per-output-channel bias, shape [out_channels].

Activation

Activation layer

Fields

name: String

Human-readable layer name (e.g. "relu1").

activation: Activation

Element-wise function applied in place; see Activation::apply.

Flatten

Reshape/Flatten layer

Fields

name: String

Human-readable layer name. Flatten collapses any input rank into a 1D tensor in [c, h, w] order.

Reshape

Reshape layer with specific dimensions

Fields

name: String

Human-readable layer name.

shape: Vec<usize>

Target tensor shape. Supported lengths are 1 (1D [size]) and 3 (3D [channels, height, width]); other lengths cause Layer::forward to return an error.

Residual

Residual/Skip connection (adds input to output)

Fields

name: String

Human-readable layer name (e.g. "resblock1").

layers: Vec<Layer>

Sub-layers applied to a clone of the input; the result is added element-wise back into the original input. Shapes of the pre- and post-sub-layer tensors must match.

Implementations

impl Layer

pub fn name(&self) -> &str

Get the layer name

pub fn forward(&self, input: &Tensor) -> Result<Tensor>

Apply this layer to input data

Trait Implementations

impl Clone for Layer

fn clone(&self) -> Layer

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Layer

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Layer

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for Layer

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.