momentum_percp.py

for i in range(0, epochs):
    mistakes = 0
    np_mistakes = 0
    np_mistakes2 = 0

    sys.stderr.write("Epoch %s...\n" % i)
    # sys.stderr.write("length of nbests is %s and first is %s\n" % (len(nbests), len(nbests[0])))
    for nbest in nbests:
        def get_sample():
            sample = []
            for j in range(0, tau):
                s1 = random.choice(nbest)
                s2 = random.choice(nbest)
                if math.fabs(s1.bleu - s2.bleu) > alpha:
                    if s1.bleu > s2.bleu:
                        sample.append((s1, s2))
                    else:
                        sample.append((s2, s1))
                else:
                    continue
            return sample

        samples = sorted(get_sample(), key=lambda h: h[0].bleu - h[1].bleu, reverse = True)[:xi]
        random.shuffle(samples)
        pre_grad = np.zeros_like(np_theta)
        
        for idx, sample in enumerate(samples):
            s1 = sample[0]
            s2 = sample[1]

            np_feat1 = np.array(s1.features, dtype=float)
            np_feat2 = np.array(s2.features, dtype=float)

            np_score1 = np.dot(np_theta2, np_feat1)
            np_score2 = np.dot(np_theta2, np_feat2)

            if np_score1 <= np_score2:
                np_mistakes2 += 1
                grad =  eta * (np_feat1 - np_feat2)
                pre_grad = pre_grad * 0.5 + grad
                np_theta2 = np_theta2 + pre_grad