Module functional 

Functional interface for neural network operations.

This module provides stateless functions that mirror the module-based activations and operations. Use these when you don’t need a module wrapper (e.g., in custom forward passes).

Example

use aprender::nn::F;
use aprender::autograd::Tensor;

let x = Tensor::randn(&[32, 10]);
let y = F::relu(&x);
let probs = F::softmax(&y, -1);

Functions

cosine_similarity_slice

Cosine similarity between two slices.

dropout

Dropout (must be called with training flag)

euclidean_distance

Euclidean distance between two slices.

gelu

GELU activation (Gaussian Error Linear Unit)

layer_norm

Layer normalization over the last dimension of an ND tensor.

leaky_relu

Leaky ReLU activation: max(negative_slope * x, x)

linear

Linear transformation: y = x @ weight^T + bias

log_softmax

Log softmax along the last dimension of an ND tensor.

log_softmax_1d

Log-softmax on a 1D slice of f32 values.

relu

ReLU activation: max(0, x)

relu_scalar

Scalar ReLU: max(0, x)

rms_norm

RMS normalization over the last dimension of an ND tensor.

sigmoid

Sigmoid activation: 1 / (1 + exp(-x))

sigmoid_scalar

Scalar sigmoid: σ(x) = 1 / (1 + exp(-x))

sigmoid_scalar_f64

Scalar sigmoid (f64): σ(x) = 1 / (1 + exp(-x))

silu

SiLU (Swish) activation: x * sigmoid(x)

silu_scalar

Scalar SiLU for non-Tensor contexts.

softmax

Softmax along the last dimension of an ND tensor.

softmax_1d

Softmax on a 1D slice of f32 values.

softmax_1d_f64

Softmax on a 1D slice of f64 values.

swiglu

SwiGLU activation: SiLU(gate) * x

swiglu_scalar

Scalar SwiGLU for non-Tensor contexts.

tanh

Tanh activation