def __init__(self, train_data, beta=1):
self.counts = Counter()
beta = int(beta)
spam = 0
total = 0
spam_prior = Counter()
ham_prior = Counter()
for row in util.get_rows(train_data):
if row['spam']:
spam += 1
del row['spam']
spam_prior.update(row)
else:
del row['spam']
ham_prior.update(row)
total += 1
ham = total - spam
prior = ((ham + beta - 1) / (total + 2*beta - 2),
(spam + beta - 1) / (total + 2*beta - 2))
for k, v in spam_prior.items():
spam_prior[k] = (v + beta -1) / (spam + 2*beta - 2)
for k, v in ham_prior.items():
ham_prior[k] = (v + beta -1) / (ham + 2*beta - 2)
self.weights = {}
self.weights['base'] = math.log(prior[1]/prior[0])
for k in set(spam_prior) | set(ham_prior):
self.weights['base'] += math.log((1 - spam_prior[k]) /
(1 - ham_prior[k]))
self.weights[k] = (math.log(spam_prior[k] / ham_prior[k]) -
math.log((1 - spam_prior[k]) /
(1 - ham_prior[k])))
def main(classifier, test_file, model_file):
classifier.dump(model_file)
count = correct = 0
for row in util.get_rows(test_file):
is_spam = row['spam']
prob = classifier.get_prob(row)
print(round(prob, 4))
if round(prob) == is_spam:
correct += 1
count += 1
print('\nAccuracy: {:.2%}'.format(correct/count))
def __init__(self, train_data, eta, sigma):
eta = float(eta)
sigma = float(sigma)
headers = ['base'] + util.get_headers(train_data)
self.weights = dict.fromkeys(headers, 0)
for i in range(100):
gradient = defaultdict(float)
for row in util.get_rows(train_data, shuffle=True):
target = row['spam']
del row['spam']
row['base'] = 1
w = target - cond_log_likelihood(self.weights, row)
for f, x in row.items():
gradient[f] = x*w - (self.weights[f] / (sigma**2))
if magnitude(gradient.values()) < 0.01:
break
for f in self.weights:
self.weights[f] += eta * gradient[f]
def __init__(self, train_data, eta, false=0):
eta = float(eta)
self.false = int(false)
headers = ['base'] + util.get_headers(train_data)
self.weights = dict.fromkeys(headers, 0)
for i in range(1000):
print(i+1)
error = False
for row in util.get_rows(train_data, false=self.false):
target = row['spam']
del row['spam']
output = self.classify(row)
if output != target:
error = True
delta = eta * (target - output)
row['base'] = 1
for x in row:
self.weights[x] += delta * row[x]
if not error:
break
