Multiclass Kernel Perceptron Library

Simple Library for Multiclass Kernel Perceptron

Very Simple Library on Multiclass Kernel Perceptron implemented with (almost) pure NumPy

Easy 5 step training sequence.

1. Import library

from Perceptron import KPerceptron
from misc import readData

2. Define hyperparameters

hparams = {'kernel': 'poly',
           'd': 4,
           'num_class': 10,
           'max_epochs': 5,
           'n_dims': 256,
           'early_stopping': False,
           'patience': 5}

hyperparameters
kernel	'poly' or 'gauss'	polynomial kernel or gaussain kernel
d	integer	if using polynomial kernel, d specify degree
c	float	if using gaussian kernel, c specify γ

Other parameters are self-explanatory.

3. Define the instance

X_train, X_test, Y_train, Y_test = readData('data/zipcombo.dat', split=True)
ker_perceptron = KPerceptron(X_train, Y_train, X_test, Y_test, hparams=hparams)

4. One line train

train_history, test_history = ker_perceptron.train()

you can get the history on training error and test error for every epoch

5. Get prediction

pred = ker_perceptron.predict(X_test)

6. Save weight for future use

ker_perceptron.save_weight('./weight/test.npy')

Warning: The data set is part of the weight. (Kernel perceptron in dual form)

If the data set is huge, the stored parameters will occupie a large space.

7. Load weight

ker_perceptron.load_weight('./weight/test.npy')

feicccccccc / Multiclass_Kernel_Perceptron