percy

Quick, straightforward, and easy-to-use Go implementations of the Perceptron and Averaged Perceptron binary classifiers.

install

While in your project directory:
$ git clone https://github.com/haydenhigg/percy

Then, import it as:

import "./percy"

use

• Regularize(arr []float64) []float64: Regularizes the L2 norm of arr to 1.
• RegularizeAll(mat [][]float64) [][]float64: Regularizes the L2 norm of every row of mat to 1.

- NewModel(weights []float64, bias float64) Model: Creates a model with input weights and an input bias.

• Train(inputs [][]float64, outputs []float64, iters int, alpha float64) Model: Trains a Perceptron classifier on the training inputs matrix and outputs vector for iters iterations with a learning rate of alpha. Returns the final model (see below).
• TrainFromModel(init Model, inputs [][]float64, outputs []float64, iters int, alpha float64) Model: The same as Train, but initializes the model to init rather than a model with weights as a zero-vector and a bias of zero.
• TrainAveraged(inputs [][]float64, outputs []float64, iters int, alpha float64) Model: The same as Train, but trains an Averaged Perceptron classifier instead.
• TrainAveragedFromModel(init Model, inputs [][]float64, outputs []float64, iters int, alpha float64) Model: The same as TrainFromWeights, but trains an Averaged Perceptron classifier instead.

• (mdl Model) Predict(x []float64) float64: Returns the predicted output (which will be {-1, 1}) for the model mdl and the input vector x.
• (mdl Model) RawPredict(x []float64) float64: Returns the output before being binarized to {-1, 1}.

Model

The Model is just a struct containing the fields Weights ([]float64) and Bias (float64).

example

package main
import (
  "fmt"
  "./percy"
)
func main() {
  inputs := [][]float64{[]float64{...}, []float64{...}, ...}
  outputs := []float64{1, -1, ...}
  iters := 200
  learningRate := 0.01
  trainedModel := percy.Train(percy.RegularizeAll(inputs), outputs, iters, learningRate)
  fmt.Println(trainedModel.Predict(percy.Regularize([]float64{...})))
}

notes

• Assumptions are not checked by this implementation. For example, if each vector of the inputs matrix does not have the same length, this will fail; if inputs and outputs are different lengths, this will fail; if learningRate is a negative number, the algorithm will not converge; etc.
• Though not necessary, it may be helpful to
  • shuffle the data before training, especially when using the standard Perceptron
  • regularize the norms of all training inputs and of all inputs to be predicted (see Regularize and RegularizeAll)
  • initialize weights to small random values rather than 0s