def ten_fold(data):
results = []
for iteration in [0]: #xrange(10):
train_set, test_set, train_labels, test_distributions=[],[],[],[]
start = iteration*(len(data)/10)
end = start + (len(data)/10)
# scale label to 100
joint_features = []
for i, item in enumerate(data):
if i >= start and i < end:
test_set.append(item[0])
test_distributions.append(item[1])
else:
# normalize label to 100 w ints for training examples
label = [int(x*100) for x in item[1]]
while sum(label) != 100:
rand = random.randint(0, len(cats) - 1)
if sum(label) < 100:
if item[1][rand] > 0.0:
label[rand] += 1
elif sum(label) > 100:
if label[rand] > 0:
label[rand] -= 1
for j, count in enumerate(label):
for k in xrange(count):
train_set.append(item[0])
train_labels.append(cats[j])
joint_features.append((item[0],cats[j]))
print >> sys.stderr, iteration, 'th iteration'
print >> sys.stderr, len(train_set), 'training examples'
# training
features = maxent.TypedMaxentFeatureEncoding.train(joint_features)
print 'features encoded'
classifier = nltk.MaxentClassifier.train(joint_features,algorithm='IIS',max_iter=2)
# testing
kl_stats = []
for f, l in itertools.izip(test_set,test_distributions):
classout = [0.]*8
probdist = classifier.prob_classify(f)
for item in probdist.samples():
classout[cats.index(item)] = probdist.prob(item)
'''
print classout
print l
print kldiv(l,classout)
print ''
'''
kl_stats.append(kldiv(l,classout))
results.append(float(sum(kl_stats))/len(kl_stats))
print results
def main():
# execute with parameters <poem-directory> <comment-directory>
dataFileName = "data.list"
if not isfile("data.list"):
print "Processing data..."
data = []
poems = PoemModel(sys.argv[1]).poems
for i, f in enumerate(sorted(os.listdir(sys.argv[2]))):
comment_f = open(sys.argv[2]+'/'+f).readlines()
poem = poems[f]
# this line removes poems from consideration if they have fewer than 10 comments - comment this out if you want to consider all poems
if len(comment_f) < 10: continue
label = get_label(comment_f)
#print label
if label == 'empty': continue
data.append((poem, label))
pickle.dump(data, open(dataFileName, "wb"))
else:
print "Loading data file..."
data = pickle.load(open(dataFileName, "rb"))
# calculate some stats
klvalues = []
for i, (_, l1) in enumerate(data):
for j, (_, l2) in enumerate(data):
klvalues.append(kldiv(l1, l2))
print "average klvalue of any pair of distributions", np.mean(klvalues)
distAverage = [sum(a) for a in zip(*[b for a,b in data])] # this isn't normalized, but it doesn't matter
klvalues = []
for i, (_, l1) in enumerate(data):
klvalues.append(kldiv(l1, distAverage))
print "average klvalue relative to average distribution", np.mean(klvalues)
ten_fold(data)