from sklearn import cross_validation as cv
from sklearn import neighbors as knn
from sklearn import ensemble as en
from sklearn import svm
from sklearn import metrics as me
import numpy as np
import itertools

# probably a simpler solution can be derived from the example in
# http://scikit-learn.org/stable/auto_examples/grid_search_digits.html#example-grid-search-digits-py


def eval_class(X,Y,cv_ext,classif,H):
    out=[]
    for TR,TE in cv_ext:
        bestacc=0.0
        bestH=None
        for hi in H:
       
            acc=0.0
            for TR2,TE2 in cv.StratifiedKFold(Y[TR],5):
                cl=train(classif,hi,X[TR2],Y[TR2])
                acc+=test(cl,X[TE2],Y[TE2])
            if acc>bestacc:
                bestacc=acc
                bestH=hi
        cl=train(classif,bestH,X[TR],Y[TR])
        out.append((test(cl,X[TE],Y[TE]),bestH))
    return out

def train(classif,hiper,X,Y):
    if classif=="svmL":
        return svm.LinearSVC(C=hiper).fit(X,Y)
    elif classif=="knn":
        return knn.KNeighborsClassifier(hiper).fit(X,Y)                        
    elif classif=="svmRBF":
        return svm.SVC(kernel="rbf",C=hiper[0],gamma=hiper[1]).fit(X,Y)
    elif classif=="rf":
        return en.RandomForestClassifier(max_features = hiper).fit(X,Y)

def test(cl,Xtest,Ytest):
    return me.accuracy_score(Ytest,cl.predict(Xtest))


with open("sonar.csv","rb") as f:
    x=[]
    y=[]
    f.readline()
    for l in f:
        z=l.strip().split(",")
        y+=z[-1][1:-1]
        z=z[:-1]
        x.append(map(float,z))

x=np.array(x)
y=np.array(y)


cv_ext=cv.StratifiedKFold(y,5)    

knnx=eval_class(x,y,cv_ext,"knn",[1,3,5,7,11,21])
print 'Acuracia media do knn =',np.mean(map(lambda x:x[0],knnx))
print 'Melhores k',map(lambda x:x[1],knnx)

rg=10**np.array(range(-3,4),dtype=np.float)

svml=eval_class(x,y,cv_ext,"svmL",rg)
print 'Acuracia media do svm Linear =',np.mean(map(lambda x:x[0],svml))
print 'Melhores C',map(lambda x:x[1],svml)

svmrbf=eval_class(x,y,cv_ext,"svmRBF",list(itertools.product(rg,rg)))
print 'Acuracia media do svm rbf =',np.mean(map(lambda x:x[0],svmrbf))
print 'Melhores C e gamma',map(lambda x:x[1],svmrbf)

rf=eval_class(x,y,cv_ext,"rf",[2,3,5,10,20,40,60])

print 'Acuracia media do rf =',np.mean(map(lambda x:x[0],rf))
print 'Melhores max_features',map(lambda x:x[1],rf)