lars.py

from modshogun import RealFeatures, RegressionLabels, LeastAngleRegression, PruneVarSubMean, PNorm
import numpy as np
from time import time
from sklearn.linear_model import LassoLars, Lars
from sklearn.metrics import mean_squared_error as mse
from sklearn import preprocessing

ran1=[3000]
ran2=[5000]#,100000]

num_feat=10
num_vec=100


def gen_data(num_feat, num_vec):
    X=np.random.rand(num_vec, num_feat)
    xtest=np.random.rand(num_vec/10, num_feat)
    
    w=np.zeros((1, num_feat))

    var=np.array(range(0, num_feat))
    var=np.random.permutation(var)

    for i in range(num_feat/3):
        w[0][var[i]]=np.random.randint(100)

    y=np.dot(w,X.T)
    ytest=np.dot(w, xtest.T)
    y=np.array(y)
    y=np.reshape(y,(num_vec,))
    ytest=np.reshape(ytest,(num_vec/10,))
    return X, y, xtest, ytest


def run():
    for f in ran1:
        for v in ran2:
            X,y, xtest, ytest =gen_data(f,v)


            y_mean = y.mean(axis=0)
            y-=y_mean
            ytest-=y_mean

            feat = RealFeatures(X.T)
            lab = RegressionLabels(y)    
            ftest = RealFeatures(xtest.T)
            ltest = RegressionLabels(ytest)

            p1=PruneVarSubMean()
            p2=PNorm(2)
            p1.init(feat)
            p1.apply_to_feature_matrix(feat)
            p1.apply_to_feature_matrix(ftest)
            p2.init(feat)
            p2.apply_to_feature_matrix(feat)
            p2.apply_to_feature_matrix(ftest)

            t1=time()
            lambda1=0.01
            modelsg = LeastAngleRegression(False)
            #modelsg.()
            modelsg.set_max_non_zero(700)
            modelsg.set_labels(lab)
            modelsg.parallel.set_num_threads(3)
            modelsg.train(feat)
            t2=time()

            timesg=t2-t1
            t3=time()
            modelsk = Lars(n_nonzero_coefs=700,precompute=False)
            modelsk.fit(X, y)
            #out = model.coef
            t4=time()
            timesk=t4-t3
            print modelsg.get_path_size()
            #print modelsg.get_w()
            print modelsk.n_iter_
            #print modelsk.coef_
            sgout=modelsg.apply(ftest)
            skout=modelsk.predict(xtest)
            
            sgout1=sgout.get_labels()
            #print sgout1, skout, ytest
            outsg=mse(ytest, sgout1)
            outsk=mse(ytest, skout)
            print "------------\n"
            print 'N:',  v, 'D:', f, 'shogun: %.6f , mse: %.6f | scikit: %.6f , mse: %.6f ' % (timesg, outsg\
                                                                                           ,timesk, outsk)

run()