Native Julia implementations of standard SVM algorithms. Currently, there are textbook style implementations of two popular linear SVM algorithms:
- Pegasos (Shalev-Schwartz et al., 2007)
- Dual Coordinate Descent (Hsieh et al., 2008)
The svm function is a wrapper for pegasos, but it is
possible to call cddual explicitly. See the source code
for the hyperparameters of the cddual function.
The demo below shows how SVMs work:
# To show how SVMs work, we'll use Fisher's iris data set
using SVM
using RDatasets
# We'll learn to separate setosa from other species
iris = dataset("datasets", "iris")
# SVM format expects observations in columns and features in rows
X = array(iris[:, 1:4])'
p, n = size(X)
# SVM format expects positive and negative examples to +1/-1
Y = [species == "setosa" ? 1.0 : -1.0 for species in iris[:Species]]
# Select a subset of the data for training, test on the rest.
train = randbool(n)
# We'll fit a model with all of the default parameters
model = svm(X[:,train], Y[train])
# And now evaluate that model on the testset
accuracy = countnz(predict(model, X[:,~train]) .== Y[~train])/countnz(~train)You may specify non-default values for the various parameters:
# The algorithm processes minibatches of data of size k
model = svm(X, Y, k = 150)
# Weight regularization is controlled by lambda
model = svm(X, Y, lambda = 0.1)
# The algorithm performs T iterations
model = svm(X, Y, T = 1000)