svm.v

module ml

import math
import rand

pub struct SVMConfig {
pub mut:
	max_iterations int = 1000
	learning_rate  f64 = 0.01
	tolerance      f64 = 1e-6
	c              f64 = 1.0 // Regularization parameter
}

pub struct DataPoint {
pub mut:
	x []f64
	y int
}

pub struct SVMModel {
pub mut:
	support_vectors []DataPoint
	alphas          []f64
	b               f64
	kernel          KernelFunction @[required]
	config          SVMConfig
}

pub struct SVM {
pub mut:
	model  &SVMModel = unsafe { nil }
	kernel KernelFunction @[required]
	config SVMConfig
}

type KernelFunction = fn ([]f64, []f64) f64

fn vector_dot(x []f64, y []f64) f64 {
	mut sum := 0.0
	for i := 0; i < x.len; i++ {
		sum += x[i] * y[i]
	}
	return sum
}

fn vector_subtract(x []f64, y []f64) []f64 {
	mut result := []f64{len: x.len}
	for i := 0; i < x.len; i++ {
		result[i] = x[i] - y[i]
	}
	return result
}

pub fn linear_kernel(x []f64, y []f64) f64 {
	return vector_dot(x, y)
}

pub fn polynomial_kernel(degree int) KernelFunction {
	return fn [degree] (x []f64, y []f64) f64 {
		return math.pow(vector_dot(x, y) + 1.0, f64(degree))
	}
}

pub fn rbf_kernel(gamma f64) KernelFunction {
	return fn [gamma] (x []f64, y []f64) f64 {
		diff := vector_subtract(x, y)
		return math.exp(-gamma * vector_dot(diff, diff))
	}
}

pub fn SVM.new(kernel KernelFunction, config SVMConfig) &SVM {
	return &SVM{
		kernel: kernel
		config: config
	}
}

pub fn (mut s SVM) train(data []DataPoint) {
	s.model = train_svm(data, s.kernel, s.config)
}

pub fn (s &SVM) predict(x []f64) int {
	return predict(s.model, x)
}

pub fn train_svm(data []DataPoint, kernel KernelFunction, config SVMConfig) &SVMModel {
	mut model := &SVMModel{
		support_vectors: []DataPoint{}
		alphas: []f64{len: data.len, init: 0.0}
		b: 0.0
		kernel: kernel
		config: config
	}

	mut passes := 0
	for {
		mut num_changed_alphas := 0
		for i in 0 .. data.len {
			ei := predict_raw(model, data[i].x) - f64(data[i].y)
			if (data[i].y * ei < -model.config.tolerance && model.alphas[i] < model.config.c)
				|| (data[i].y * ei > model.config.tolerance && model.alphas[i] > 0) {
				j := rand.int_in_range(0, data.len - 1) or { panic(err) }
				ej := predict_raw(model, data[j].x) - f64(data[j].y)

				alpha_i_old := model.alphas[i]
				alpha_j_old := model.alphas[j]

				mut l, mut h := 0.0, 0.0
				if data[i].y != data[j].y {
					l = math.max(0.0, model.alphas[j] - model.alphas[i])
					h = math.min(model.config.c, model.config.c + model.alphas[j] - model.alphas[i])
				} else {
					l = math.max(0.0, model.alphas[i] + model.alphas[j] - model.config.c)
					h = math.min(model.config.c, model.alphas[i] + model.alphas[j])
				}

				if l == h {
					continue
				}

				eta := 2 * model.kernel(data[i].x, data[j].x) - model.kernel(data[i].x,
					data[i].x) - model.kernel(data[j].x, data[j].x)

				if eta >= 0 {
					continue
				}

				model.alphas[j] = alpha_j_old - f64(data[j].y) * (ei - ej) / eta
				model.alphas[j] = math.max(l, math.min(h, model.alphas[j]))

				if math.abs(model.alphas[j] - alpha_j_old) < 1e-5 {
					continue
				}

				model.alphas[i] = alpha_i_old +
					f64(data[i].y * data[j].y) * (alpha_j_old - model.alphas[j])

				b1 := model.b - ei - f64(data[i].y) * (model.alphas[i] - alpha_i_old) * model.kernel(data[i].x,
					data[i].x) - f64(data[j].y) * (model.alphas[j] - alpha_j_old) * model.kernel(data[i].x,
					data[j].x)

				b2 := model.b - ej - f64(data[i].y) * (model.alphas[i] - alpha_i_old) * model.kernel(data[i].x,
					data[j].x) - f64(data[j].y) * (model.alphas[j] - alpha_j_old) * model.kernel(data[j].x,
					data[j].x)

				if 0 < model.alphas[i] && model.alphas[i] < model.config.c {
					model.b = b1
				} else if 0 < model.alphas[j] && model.alphas[j] < model.config.c {
					model.b = b2
				} else {
					model.b = (b1 + b2) / 2
				}

				num_changed_alphas++
			}
		}

		if num_changed_alphas == 0 {
			passes++
		} else {
			passes = 0
		}

		if passes >= model.config.max_iterations {
			break
		}
	}

	for i in 0 .. data.len {
		if model.alphas[i] > 0 {
			model.support_vectors << data[i]
		}
	}

	return model
}

fn predict_raw(model &SVMModel, x []f64) f64 {
	mut sum := 0.0
	for i, sv in model.support_vectors {
		sum += model.alphas[i] * f64(sv.y) * model.kernel(x, sv.x)
	}
	return sum + model.b
}

pub fn predict(model &SVMModel, x []f64) int {
	return if predict_raw(model, x) >= 0 { 1 } else { -1 }
}