pyphmm.py

#!/usr/bin/env python
import sys
import random
import numpy as np

from eta import ETA

import pyximport; pyximport.install()
import polya
from polya import Polya, PolyaHyper, Uniform, HMM, LM

from util import *
from logarithmetic import *

from hpylm import hpylm

def pyphmm(kind, text, words, S, M=2, L=None):
    trans = hpylm(Uniform(S), M+1)
    Eh = PolyaHyper('HMM-E')
    Uword = Uniform(len(words)); Uchar = Uniform(0x100)
    
    if L is None: # no language model
        emit = [Polya(Eh, Uword) for _ in range(S)]
    else:
        lms = [LM(hpylm(Uchar, L+1), L, words) for s in range(S)]
        emit = [Polya(Eh, lms[s]) for s in range(S)]
    
    hmm = HMM(kind, text, trans, emit, M, S)
    hmm.train()
    return hmm

def main(n_iter=50):
    random.seed(0); np.random.seed(0)
    text_vocab = DynEnum(); tag_vocab = DynEnum()
    text = []; gold_tags = []
    for line in sys.stdin:
        word,tag = line.split()
        text.append(text_vocab(word))
        gold_tags.append(tag_vocab(tag))
    
    hmm = pyphmm(1, text, text_vocab.names, len(tag_vocab), L=1)
    
    eta = ETA(n_iter); eta.print_status(0, extra='starting...')
    for i in range(n_iter):
        print 'iteration', i
        polya.resample()
        post = hmm.resample()
        status = "POST %.2fb | M-1 %.2f%%" \
            % (-log(post,2) / len(text), hmm.M1(gold_tags) * 100)
        print 'status:', status
        eta.print_status(i+1, extra=status)
    eta.done()
    print>>sys.stderr, 'M-1:', hmm.M1(gold_tags) * 100
if __name__ == '__main__': main()