train.py

#!/usr/bin/env python3

from __future__ import division # no need for python3, but just in case used w/ python2

import sys
import time
from svector import svector

def read_from(textfile):
    for line in open(textfile):
        label, words = line.strip().split("\t")
        yield (1 if label=="+" else -1, words.split())

def make_vector(words):
    v = svector()
    #add bias
    v[0]=1
    for word in words:
        v[word] += 1
    return v
    
def test(devfile, model):
    tot, err = 0, 0
    for i, (label, words) in enumerate(read_from(devfile), 1): # note 1...|D|
        err += label * (model.dot(make_vector(words))) <= 0
    return err/i  # i is |D| now
            
def train(trainfile, devfile, epochs=2):
    t = time.time()
    best_err = 1.
    model = svector()
    c = 0
    wa = svector()
    for it in range(1, epochs+1):
        updates = 0
        for i, (label, words) in enumerate(read_from(trainfile), 1): # label is +1 or -1
            sent = make_vector(words)
            if label * (model.dot(sent)) <= 0:
                updates += 1
                model += label * sent
                wa += c*label*sent
                model = c*model - wa
            c += 1
        dev_err = test(devfile, model)
        best_err = min(best_err, dev_err)
        print("epoch %d, update %.1f%%, dev %.1f%%" % (it, updates / i * 100, dev_err * 100))
    print("best dev err %.1f%%, |w|=%d, time: %.1f secs" % (best_err * 100, len(model), time.time() - t))

if __name__ == "__main__":
    train(sys.argv[1], sys.argv[2], 1)