rustframe/src/compute/models/dense_nn.rs

552 lines
18 KiB
Rust
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

//! A minimal dense neural network implementation for educational purposes.
//!
//! Layers operate on [`Matrix`] values and support ReLU and Sigmoid
//! activations. This is not meant to be a performant deeplearning framework
//! but rather a small example of how the surrounding matrix utilities can be
//! composed.
//!
//! ```
//! use rustframe::compute::models::dense_nn::{ActivationKind, DenseNN, DenseNNConfig, InitializerKind, LossKind};
//! use rustframe::matrix::Matrix;
//!
//! // Tiny network with one input and one output neuron.
//! let config = DenseNNConfig {
//! input_size: 1,
//! hidden_layers: vec![],
//! output_size: 1,
//! activations: vec![ActivationKind::Relu],
//! initializer: InitializerKind::Uniform(0.5),
//! loss: LossKind::MSE,
//! learning_rate: 0.1,
//! epochs: 1,
//! };
//! let mut nn = DenseNN::new(config);
//! let x = Matrix::from_vec(vec![1.0, 2.0], 2, 1);
//! let y = Matrix::from_vec(vec![2.0, 3.0], 2, 1);
//! nn.train(&x, &y);
//! ```
use crate::compute::models::activations::{drelu, relu, sigmoid};
use crate::matrix::{Matrix, SeriesOps};
use crate::random::prelude::*;
/// Supported activation functions
#[derive(Clone)]
pub enum ActivationKind {
Relu,
Sigmoid,
Tanh,
}
impl ActivationKind {
/// Apply activation elementwise
pub fn forward(&self, z: &Matrix<f64>) -> Matrix<f64> {
match self {
ActivationKind::Relu => relu(z),
ActivationKind::Sigmoid => sigmoid(z),
ActivationKind::Tanh => z.map(|v| v.tanh()),
}
}
/// Compute elementwise derivative w.r.t. pre-activation z
pub fn derivative(&self, z: &Matrix<f64>) -> Matrix<f64> {
match self {
ActivationKind::Relu => drelu(z),
ActivationKind::Sigmoid => {
let s = sigmoid(z);
s.zip(&s, |si, sj| si * (1.0 - sj))
}
ActivationKind::Tanh => z.map(|v| 1.0 - v.tanh().powi(2)),
}
}
}
/// Weight initialization schemes
#[derive(Clone)]
pub enum InitializerKind {
/// Uniform(-limit .. limit)
Uniform(f64),
/// Xavier/Glorot uniform
Xavier,
/// He (Kaiming) uniform
He,
}
impl InitializerKind {
pub fn initialize(&self, rows: usize, cols: usize) -> Matrix<f64> {
let mut rng = rng();
let fan_in = rows;
let fan_out = cols;
let limit = match self {
InitializerKind::Uniform(l) => *l,
InitializerKind::Xavier => (6.0 / (fan_in + fan_out) as f64).sqrt(),
InitializerKind::He => (2.0 / fan_in as f64).sqrt(),
};
let data = (0..rows * cols)
.map(|_| rng.random_range(-limit..limit))
.collect::<Vec<_>>();
Matrix::from_vec(data, rows, cols)
}
}
/// Supported losses
#[derive(Clone)]
pub enum LossKind {
/// Mean Squared Error: L = 1/m * sum((y_hat - y)^2)
MSE,
/// Binary Cross-Entropy: L = -1/m * sum(y*log(y_hat) + (1-y)*log(1-y_hat))
BCE,
}
impl LossKind {
/// Compute gradient dL/dy_hat (before applying activation derivative)
pub fn gradient(&self, y_hat: &Matrix<f64>, y: &Matrix<f64>) -> Matrix<f64> {
let m = y.rows() as f64;
match self {
LossKind::MSE => (y_hat - y) * (2.0 / m),
LossKind::BCE => (y_hat - y) * (1.0 / m),
}
}
}
/// Configuration for a dense neural network
pub struct DenseNNConfig {
pub input_size: usize,
pub hidden_layers: Vec<usize>,
/// Must have length = hidden_layers.len() + 1
pub activations: Vec<ActivationKind>,
pub output_size: usize,
pub initializer: InitializerKind,
pub loss: LossKind,
pub learning_rate: f64,
pub epochs: usize,
}
/// A multi-layer perceptron with full configurability
pub struct DenseNN {
weights: Vec<Matrix<f64>>,
biases: Vec<Matrix<f64>>,
activations: Vec<ActivationKind>,
loss: LossKind,
lr: f64,
epochs: usize,
}
impl DenseNN {
/// Build a new DenseNN from the given configuration
pub fn new(config: DenseNNConfig) -> Self {
let mut sizes = vec![config.input_size];
sizes.extend(&config.hidden_layers);
sizes.push(config.output_size);
assert_eq!(
config.activations.len(),
sizes.len() - 1,
"Number of activation functions must match number of layers"
);
let mut weights = Vec::with_capacity(sizes.len() - 1);
let mut biases = Vec::with_capacity(sizes.len() - 1);
for i in 0..sizes.len() - 1 {
let w = config.initializer.initialize(sizes[i], sizes[i + 1]);
let b = Matrix::zeros(1, sizes[i + 1]);
weights.push(w);
biases.push(b);
}
DenseNN {
weights,
biases,
activations: config.activations,
loss: config.loss,
lr: config.learning_rate,
epochs: config.epochs,
}
}
/// Perform a full forward pass, returning pre-activations (z) and activations (a)
fn forward_full(&self, x: &Matrix<f64>) -> (Vec<Matrix<f64>>, Vec<Matrix<f64>>) {
let mut zs = Vec::with_capacity(self.weights.len());
let mut activs = Vec::with_capacity(self.weights.len() + 1);
activs.push(x.clone());
let mut a = x.clone();
for (i, (w, b)) in self.weights.iter().zip(self.biases.iter()).enumerate() {
let z = &a.dot(w) + &Matrix::repeat_rows(b, a.rows());
let a_next = self.activations[i].forward(&z);
zs.push(z);
activs.push(a_next.clone());
a = a_next;
}
(zs, activs)
}
/// Train the network on inputs X and targets Y
pub fn train(&mut self, x: &Matrix<f64>, y: &Matrix<f64>) {
let m = x.rows() as f64;
for _ in 0..self.epochs {
let (zs, activs) = self.forward_full(x);
let y_hat = activs.last().unwrap().clone();
// Initial delta (dL/dz) on output
let mut delta = match self.loss {
LossKind::BCE => self.loss.gradient(&y_hat, y),
LossKind::MSE => {
let grad = self.loss.gradient(&y_hat, y);
let dz = self
.activations
.last()
.unwrap()
.derivative(zs.last().unwrap());
grad.zip(&dz, |g, da| g * da)
}
};
// Backpropagate through layers
for l in (0..self.weights.len()).rev() {
let a_prev = &activs[l];
let dw = a_prev.transpose().dot(&delta) / m;
let db = Matrix::from_vec(delta.sum_vertical(), 1, delta.cols()) / m;
// Update weights & biases
self.weights[l] = &self.weights[l] - &(dw * self.lr);
self.biases[l] = &self.biases[l] - &(db * self.lr);
// Propagate delta to previous layer
if l > 0 {
let w_t = self.weights[l].transpose();
let da = self.activations[l - 1].derivative(&zs[l - 1]);
delta = delta.dot(&w_t).zip(&da, |d, a| d * a);
}
}
}
}
/// Run a forward pass and return the network's output
pub fn predict(&self, x: &Matrix<f64>) -> Matrix<f64> {
let mut a = x.clone();
for (i, (w, b)) in self.weights.iter().zip(self.biases.iter()).enumerate() {
let z = &a.dot(w) + &Matrix::repeat_rows(b, a.rows());
a = self.activations[i].forward(&z);
}
a
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::matrix::Matrix;
/// Compute MSE = 1/m * Σ (ŷ - y)²
fn mse_loss(y_hat: &Matrix<f64>, y: &Matrix<f64>) -> f64 {
let m = y.rows() as f64;
y_hat
.zip(y, |yh, yv| (yh - yv).powi(2))
.data()
.iter()
.sum::<f64>()
/ m
}
#[test]
fn test_predict_shape() {
let config = DenseNNConfig {
input_size: 1,
hidden_layers: vec![2],
activations: vec![ActivationKind::Relu, ActivationKind::Sigmoid],
output_size: 1,
initializer: InitializerKind::Uniform(0.1),
loss: LossKind::MSE,
learning_rate: 0.01,
epochs: 0,
};
let model = DenseNN::new(config);
let x = Matrix::from_vec(vec![1.0, 2.0, 3.0], 3, 1);
let preds = model.predict(&x);
assert_eq!(preds.rows(), 3);
assert_eq!(preds.cols(), 1);
}
#[test]
#[should_panic(expected = "Number of activation functions must match number of layers")]
fn test_invalid_activation_count() {
let config = DenseNNConfig {
input_size: 2,
hidden_layers: vec![3],
activations: vec![ActivationKind::Relu], // Only one activation for two layers
output_size: 1,
initializer: InitializerKind::Uniform(0.1),
loss: LossKind::MSE,
learning_rate: 0.01,
epochs: 0,
};
let _model = DenseNN::new(config);
}
#[test]
fn test_train_no_epochs_does_nothing() {
let config = DenseNNConfig {
input_size: 1,
hidden_layers: vec![2],
activations: vec![ActivationKind::Relu, ActivationKind::Sigmoid],
output_size: 1,
initializer: InitializerKind::Uniform(0.1),
loss: LossKind::MSE,
learning_rate: 0.01,
epochs: 0,
};
let mut model = DenseNN::new(config);
let x = Matrix::from_vec(vec![0.0, 1.0], 2, 1);
let y = Matrix::from_vec(vec![0.0, 1.0], 2, 1);
let before = model.predict(&x);
model.train(&x, &y);
let after = model.predict(&x);
for i in 0..before.rows() {
for j in 0..before.cols() {
// "prediction changed despite 0 epochs"
assert!((before[(i, j)] - after[(i, j)]).abs() < 1e-12);
}
}
}
#[test]
fn test_train_one_epoch_changes_predictions() {
// Single-layer sigmoid regression so gradients flow.
let config = DenseNNConfig {
input_size: 1,
hidden_layers: vec![],
activations: vec![ActivationKind::Sigmoid],
output_size: 1,
initializer: InitializerKind::Uniform(0.1),
loss: LossKind::MSE,
learning_rate: 1.0,
epochs: 1,
};
let mut model = DenseNN::new(config);
let x = Matrix::from_vec(vec![0.0, 1.0], 2, 1);
let y = Matrix::from_vec(vec![0.0, 1.0], 2, 1);
let before = model.predict(&x);
model.train(&x, &y);
let after = model.predict(&x);
// At least one of the two outputs must move by >ϵ
let mut moved = false;
for i in 0..before.rows() {
if (before[(i, 0)] - after[(i, 0)]).abs() > 1e-8 {
moved = true;
}
}
assert!(moved, "predictions did not change after 1 epoch");
}
#[test]
fn test_training_reduces_mse_loss() {
// Same singlelayer sigmoid setup; check loss goes down.
let config = DenseNNConfig {
input_size: 1,
hidden_layers: vec![],
activations: vec![ActivationKind::Sigmoid],
output_size: 1,
initializer: InitializerKind::Uniform(0.1),
loss: LossKind::MSE,
learning_rate: 1.0,
epochs: 10,
};
let mut model = DenseNN::new(config);
let x = Matrix::from_vec(vec![0.0, 1.0, 0.5], 3, 1);
let y = Matrix::from_vec(vec![0.0, 1.0, 0.5], 3, 1);
let before_preds = model.predict(&x);
let before_loss = mse_loss(&before_preds, &y);
model.train(&x, &y);
let after_preds = model.predict(&x);
let after_loss = mse_loss(&after_preds, &y);
// MSE did not decrease (before: {}, after: {})
assert!(after_loss < before_loss);
}
#[test]
fn test_activation_kind_forward_tanh() {
let input = Matrix::from_vec(vec![-1.0, 0.0, 1.0], 3, 1);
let expected = Matrix::from_vec(vec![-0.76159415595, 0.0, 0.76159415595], 3, 1);
let output = ActivationKind::Tanh.forward(&input);
for i in 0..input.rows() {
for j in 0..input.cols() {
// Tanh forward output mismatch at ({}, {})
assert!((output[(i, j)] - expected[(i, j)]).abs() < 1e-9);
}
}
}
#[test]
fn test_activation_kind_derivative_relu() {
let input = Matrix::from_vec(vec![-1.0, 0.0, 1.0], 3, 1);
let expected = Matrix::from_vec(vec![0.0, 0.0, 1.0], 3, 1);
let output = ActivationKind::Relu.derivative(&input);
for i in 0..input.rows() {
for j in 0..input.cols() {
// "ReLU derivative output mismatch at ({}, {})"
assert!((output[(i, j)] - expected[(i, j)]).abs() < 1e-9);
}
}
}
#[test]
fn test_activation_kind_derivative_tanh() {
let input = Matrix::from_vec(vec![-1.0, 0.0, 1.0], 3, 1);
let expected = Matrix::from_vec(vec![0.41997434161, 1.0, 0.41997434161], 3, 1); // 1 - tanh(x)^2
let output = ActivationKind::Tanh.derivative(&input);
for i in 0..input.rows() {
for j in 0..input.cols() {
// "Tanh derivative output mismatch at ({}, {})"
assert!((output[(i, j)] - expected[(i, j)]).abs() < 1e-9);
}
}
}
#[test]
fn test_initializer_kind_xavier() {
let rows = 10;
let cols = 20;
let initializer = InitializerKind::Xavier;
let matrix = initializer.initialize(rows, cols);
let limit = (6.0 / (rows + cols) as f64).sqrt();
assert_eq!(matrix.rows(), rows);
assert_eq!(matrix.cols(), cols);
for val in matrix.data() {
// Xavier initialized value out of range
assert!(*val >= -limit && *val <= limit);
}
}
#[test]
fn test_initializer_kind_he() {
let rows = 10;
let cols = 20;
let initializer = InitializerKind::He;
let matrix = initializer.initialize(rows, cols);
let limit = (2.0 / rows as f64).sqrt();
assert_eq!(matrix.rows(), rows);
assert_eq!(matrix.cols(), cols);
for val in matrix.data() {
// He initialized value out of range
assert!(*val >= -limit && *val <= limit);
}
}
#[test]
fn test_loss_kind_bce_gradient() {
let y_hat = Matrix::from_vec(vec![0.1, 0.9, 0.4], 3, 1);
let y = Matrix::from_vec(vec![0.0, 1.0, 0.5], 3, 1);
let expected_gradient = Matrix::from_vec(vec![0.1 / 3.0, -0.1 / 3.0, -0.1 / 3.0], 3, 1); // (y_hat - y) * (1.0 / m)
let output_gradient = LossKind::BCE.gradient(&y_hat, &y);
assert_eq!(output_gradient.rows(), expected_gradient.rows());
assert_eq!(output_gradient.cols(), expected_gradient.cols());
for i in 0..output_gradient.rows() {
for j in 0..output_gradient.cols() {
// BCE gradient output mismatch at ({}, {})
assert!((output_gradient[(i, j)] - expected_gradient[(i, j)]).abs() < 1e-9);
}
}
}
#[test]
fn test_training_reduces_bce_loss() {
// Single-layer sigmoid setup; check BCE loss goes down.
let config = DenseNNConfig {
input_size: 1,
hidden_layers: vec![],
activations: vec![ActivationKind::Sigmoid],
output_size: 1,
initializer: InitializerKind::Uniform(0.1),
loss: LossKind::BCE,
learning_rate: 1.0,
epochs: 10,
};
let mut model = DenseNN::new(config);
let x = Matrix::from_vec(vec![0.0, 1.0, 0.5], 3, 1);
let y = Matrix::from_vec(vec![0.0, 1.0, 0.5], 3, 1);
let before_preds = model.predict(&x);
// BCE loss calculation for testing
let before_loss = -1.0 / (y.rows() as f64)
* before_preds
.zip(&y, |yh, yv| yv * yh.ln() + (1.0 - yv) * (1.0 - yh).ln())
.data()
.iter()
.sum::<f64>();
model.train(&x, &y);
let after_preds = model.predict(&x);
let after_loss = -1.0 / (y.rows() as f64)
* after_preds
.zip(&y, |yh, yv| yv * yh.ln() + (1.0 - yv) * (1.0 - yh).ln())
.data()
.iter()
.sum::<f64>();
// BCE did not decrease (before: {}, after: {})
assert!(after_loss < before_loss,);
}
#[test]
fn test_train_backprop_delta_propagation() {
// Network with two layers to test delta propagation to previous layer (l > 0)
let config = DenseNNConfig {
input_size: 2,
hidden_layers: vec![3],
activations: vec![ActivationKind::Sigmoid, ActivationKind::Sigmoid],
output_size: 1,
initializer: InitializerKind::Uniform(0.1),
loss: LossKind::MSE,
learning_rate: 0.1,
epochs: 1,
};
let mut model = DenseNN::new(config);
// Store initial weights and biases to compare after training
let initial_weights_l0 = model.weights[0].clone();
let initial_biases_l0 = model.biases[0].clone();
let initial_weights_l1 = model.weights[1].clone();
let initial_biases_l1 = model.biases[1].clone();
let x = Matrix::from_vec(vec![0.1, 0.2, 0.3, 0.4], 2, 2);
let y = Matrix::from_vec(vec![0.5, 0.6], 2, 1);
model.train(&x, &y);
// Verify that weights and biases of both layers have changed,
// implying delta propagation occurred for l > 0
// Weights of first layer did not change, delta propagation might not have occurred
assert!(model.weights[0] != initial_weights_l0);
// Biases of first layer did not change, delta propagation might not have occurred
assert!(model.biases[0] != initial_biases_l0);
// Weights of second layer did not change
assert!(model.weights[1] != initial_weights_l1);
// Biases of second layer did not change
assert!(model.biases[1] != initial_biases_l1);
}
}