A kotlin port of https://github.com/karpathy/micrograd.
See OptimizationTest.kt for how to use it:
kt-micrograd/src/test/kotlin/OptimizationTest.kt
Lines 4 to 45 in 3f33b2b
| @Test fun `optimization test`() { | |
| val model = MLP(3, listOf(4, 4, 1)) | |
| val inputs = listOf( | |
| listOf(2.0, 3.0, -1.0), | |
| listOf(3.0, -1.0, 0.5), | |
| listOf(0.5, 1.0, -1.0), | |
| listOf(1.0, 1.0, -1.0), | |
| ) | |
| // desired targets | |
| val ys = listOf(1.0, -1.0, -1.0, 1.0) | |
| println("number of parameters: " + model.parameters().size) | |
| fun loss(): Pair<Value, List<Value>> { | |
| // Assuming `model` is a function that takes a list of `Value` and returns a `Value` | |
| val scores = inputs.flatMap { model(it) } | |
| val loss = ys.zip(scores) | |
| .map { (ygt, yout) -> (yout - ygt).pow(2.0) } | |
| .reduce { acc, value -> acc + value } | |
| return loss to scores | |
| } | |
| // gradient descent | |
| val maxSteps = 1000 | |
| for (k in 0..maxSteps) { | |
| // forward | |
| val (totalLoss, predictions) = loss() | |
| // backward | |
| model.zeroGrad() | |
| totalLoss.backward() | |
| // update(sgd) | |
| val learningRate = 0.02 | |
| for (p in model.parameters()) { | |
| p.data -= learningRate * p.grad | |
| } | |
| if (k % 100 == 0) { | |
| println("step $k loss ${totalLoss.data}, learningRate=$learningRate") | |
| } | |
| if (k == maxSteps) { | |
| println("predictions: ${predictions.map { it.data }}") | |
| } | |
| } | |
| } |