def splitonmaxarg(self, x_tl, y_tl, features, D_t, isSparse=0):
ret = []
pplus = sum(D_t * (y_tl > 0))
for feature_i in range(features):
(dv, err) = ds.stump_fit(x_tl[:, feature_i], y_tl, D_t, pplus)
ret.append((feature_i, dv, err))
a_ret = array(ret)
arg = argmax(abs(0.5 - a_ret[:, 2]))
return a_ret[arg]
def adaboost_train (x, y, T):
cf = x.shape[1]
n = y.shape[0]
weights = ones(n)/n
H = []
A = []
I = []
TE = []
for t in range(T):
pplus = sum(weights * (y > 0))
# Let's train on all the features and find the one that works the best
decisionVariables = []
score = []
we = []
for idx in range(cf):
f = x[:,idx]
# train the stump
(dv, err) = ds.stump_fit(f, y, weights, pplus)
we.append( err )
decisionVariables.append(dv)
# score the classifiers on all features for this round
score.append(abs(.5-err))
print "Round: ", t, str(datetime.now())
# choose the one feature we'll use for this round's classifier
I.append(np.argmax(score))
H.append(decisionVariables[I[t]])
eps = we[I[t]]
# calculate our alpha
A.append(.5 * math.log((1-eps)/eps))
# update the weights
numerators = weights * np.exp( -A[t] * y * ds.stump_predict(x[:,I[t]], H[t]) )
Z = numerators.sum()
weights = numerators / Z
# Calculate the overall training errors
y_hat = adaboost_predict(A,H,I,x, len(A))
TE.append((y_hat * y < 0).sum() / float(n))
return A, H, I, TE
#!/usr/bin/python2 from scipy import * import scipy.sparse as sp import dstump as ds (f, y) = ds.two_clusters(100) pr = ones(len(y))/len(y) #This quantity is invariant for each Adaboost step, and helps us take #advantage of sparsity. pplus = sum(pr * (y > 0)) #The decision stump training routine accepts either a dense 1-d #array or a sparse 1-d CSC matrix. The resulting decision variable #might be different for dense and sparse data, but the errors are #the same. See implementation for details. (dv, err) = ds.stump_fit(f, y, pr, pplus) #Inplace transpose of a CSR matrix gives a CSC matrix. fs = sp.csr_matrix(f).T (dvs, errs) = ds.stump_fit(fs, y, pr, pplus)
