def yhat(x, w):
"""
Get predicted labels under the current hypothesis
:param w: Hypothesis to be evaluated
:param x: Feature matrix
:return: List of predicted labels
"""
return list(map(lambda x_i: PLA.sign(x_i, w), x))
def q17(x, y):
# Question 17: Shuffled input, update step = 0.5, 2000 trials
print('Question 17: Running PLA on training samples with update step 0.5 and shuffled input ...')
updates = []
for i in range(2000):
updates.append(PLA.pla(x, y, True, 0.5))
print('Average number of updates for PLA to converge with update step 0.5 and shuffled input {}'
.format(sum(updates) / len(updates)))
print()
def hw15():
f = open('./hw1_15_train.dat', 'r')
data = []
for line in f:
l = line.split()
data.append([[float(i) for i in l[:-1]], int(l[-1])])
f.close()
print PLA(data)
def hw16():
f = open('./hw1_15_train.dat', 'r')
data = []
for line in f:
l = line.split()
data.append([[float(i) for i in l[:-1]], int(l[-1])])
f.close()
updates = []
for i in range(0, 2000):
w, u = PLA(data, True)
print '#%d' % (i), w, u
updates.append(u)
print 'Avg:', float(sum(updates)) / len(updates)
err_idx = np.where(yhat != y)[0]
if len(err_idx) == 0:
break
# randomly choose a sample where the current weight makes a mistake
i = err_idx[np.random.permutation(len(err_idx))[0]]
# adjust weight vector based on the chosen sample
w += step * y[i, 0] * np.matrix(x[i]).T
# get errors after weight update
yhat = Pocket.yhat(x, w)
new_err = Pocket.err_rate(yhat, y)
if new_err < curr_err:
w_pocket = w.copy()
curr_err = new_err
return w_pocket, w
# for local testing only
if __name__ == '__main__':
x_train, y_train = PLA.load_samples('hw1_18_train.dat')
x_eval, y_eval = PLA.load_samples('hw1_18_test.dat')
w_pocket, w = Pocket.pocket(x_train, y_train, 1, 50)
err_rate = Pocket.err_rate(Pocket.yhat(x_eval, w_pocket), y_eval)
print('OTS error rate for w_pocket: {}'.format(err_rate))
def q15(x, y):
# Question 15: No input shuffling, update step = 1
print('Question 15: Running PLA on training samples with update step 1 and no shuffling ...')
print('Number of updates to PLA convergence: {}'.format(PLA.pla(x, y, False, 1)))
print()