Float NDArray library accelerated with Accelerate Framework
func testGradientDescent() {
// y = 0.3*x^2 + 0.2*x + 0.1
let start = Date()
let xs = NDArray.linspace(low: -1, high: 1, count: 300)
// data
var ys = 0.3*xs*xs + 0.2*xs + 0.1
ys += NDArray.normal(mu: 0, sigma: 0.03, shape: xs.shape)
print("xs: \(xs.shape), ys: \(ys.shape)")
// x^2, x^1, x^0
let features = NDArray.stack([xs*xs, xs, NDArray.ones(xs.shape)], newAxis: -1)
print("features: \(features.shape)")
var theta = NDArray([1, 1, 1])
let alpha: Float = 0.1
for i in 0..<2000 {
// calculate loss
let distance = sum(theta * features, along: 1) - ys
let loss = mean(distance**2, along: 0) / 2
// Update parameters
let grads = distance.reshaped([-1, 1]) * features
let update = alpha * mean(grads, along: 0)
theta -= update
if i%100 == 0 {
print("\nstep: \(i)")
print("loss: \(loss.asScalar())")
print("grads: \(grads.shape)")
print("update: \(update)")
print("theta: \(theta)")
}
}
print("\nanswer")
print("theta: \(theta)")
let distance = sum(theta * features, along: 1) - ys
let loss = mean(distance**2, along: 0) / 2
print("loss: \(loss.asScalar())")
print("elapsed time: \(Date().timeIntervalSince(start))sec")
print("")
}answer
theta: NDArray(shape: [3], data: [0.300824702, 0.199845955, 0.102966547], strides: [1], baseOffset: 0)
loss: 0.00041115
elapsed time: 0.166108965873718sec s