2013/09/16「嵐のPRMLハッカソン」で書いてたコード

latin.dat

amō    indicative	active	present	sg	1
amās	indicative	active	present	sg	2
amat	indicative	active	present	sg	3
amāmus	indicative	active	present	pl	1
amātis	indicative	active	present	pl	2
amant	indicative	active	present	pl	3
amābam	indicative	active	imperfect	sg	1
amābās	indicative	active	imperfect	sg	2
amābat	indicative	active	imperfect	sg	3
amābāmus	indicative	active	imperfect	pl	1
amābātis	indicative	active	imperfect	pl	2
amābant	indicative	active	imperfect	pl	3
amābō	indicative	active	future	sg	1
amābis	indicative	active	future	sg	2
amābit	indicative	active	future	sg	3
amābimus	indicative	active	future	pl	1
amābitis	indicative	active	future	pl	2
amābunt	indicative	active	future	pl	3
amāvī	indicative	active	perfect	sg	1
amāvistī	indicative	active	perfect	sg	2
amāvit	indicative	active	perfect	sg	3
amāvimus	indicative	active	perfect	pl	1
amāvistis	indicative	active	perfect	pl	2
amāvēre	indicative	active	perfect	pl	3
amāveram	indicative	active	past-perfect	sg	1
amāverās	indicative	active	past-perfect	sg	2
amāverat	indicative	active	past-perfect	sg	3
amāverāmus	indicative	active	past-perfect	pl	1
amāverātis	indicative	active	past-perfect	pl	2
amāverant	indicative	active	past-perfect	pl	3
amāverō	indicative	active	future-perfect	sg	1
amāveris	indicative	active	future-perfect	sg	2
amāverit	indicative	active	future-perfect	sg	3
amāverimus	indicative	active	future-perfect	pl	1
amāveritis	indicative	active	future-perfect	pl	2
amāverint	indicative	active	future-perfect	pl	3

latin_loader.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
import sys

IGNORE_CASE = True
VALID_TAIL_LENGTH = 10

from itertools import chain
def flatten_list(l):
    return list(chain.from_iterable(l))

import numpy as np

# a b c d e f g h ij k l  m  n  o  p  q  r  s  t  uv[w] x  y  z
# 0 1 2 3 4 5 6 7 8  9 10 11 12 13 14 15 16 17 18 19    20 21 22
def trans(ch):
    def conv(i):
        if i <= 7: return i
        elif i in (8, 9): return 8
        elif i <= 19: return i - 1
        elif i in (20, 21): return 19
        elif i <= 25: return i - 3
        else: return -1

    o = ord(ch)

    if o in (256, 257): # a^
        a, m, c = 0, True, o == 256
    elif o in (274, 275): # e^
        a, m, c = 4, True, o == 274
    elif o in (298, 299): # i^
        a, m, c = 8, True, o == 298
    elif o in (332, 333): # o^
        a, m, c = 13, True, o == 332
    elif o in (362, 363): # u^
        a, m, c = 19, True, o == 362
    elif 65 <= o <= 90:
        a, m, c = conv(o-65), False, True
    elif 97 <= o <= 122:
        a, m, c = conv(o-97), False, False
    else:
        a, m, c = None, False, False

    # return a, m, c
    if IGNORE_CASE:
        vec = [0]*24
        if a: vec[a] = 1
        if m: vec[23] = 1
    else:
        vec = [0]*25
        if a: vec[a] = 1
        if m: vec[23] = 1
        if c: vec[24] = 1
    return vec


def vectorize_surface(surface):
    surface = (" " * VALID_TAIL_LENGTH) + surface
    tail = surface[:-VALID_TAIL_LENGTH-1:-1] # 後ろから10文字
    return map(trans, tail)


def vectorize_mood(mood): # [indicative, ...]
    vec = [0, 0, 0]
    if mood == 'indicative': vec[0] = 1
    return vec

def vectorize_voice(voice):
    vec = [0, 0]
    if voice == 'active': vec[0] = 1
    if voice == 'passive': vec[1] = 1
    return vec

def vectorize_tense(tense):
    vec = [0, 0, 0, 0, 0, 0]
    if tense == 'present': vec[0] = 1
    if tense == 'imperfect': vec[1] = 1
    if tense == 'future': vec[2] = 1
    if tense == 'perfect': vec[3] = 1
    if tense == 'past-perfect': vec[4] = 1
    if tense == 'future-perfect': vec[5] = 1
    return vec

def vectorize_number(number):
    vec = [0, 0]
    if number == 'sg': vec[0] = 1
    if number == 'pl': vec[1] = 1
    return vec

def vectorize_person(person):
    vec = [0, 0, 0]
    if 1 <= person <= 3:
        vec[person] = 1
    return vec

def load_data(data_filepath):
    xs = []
    ys = []

    with open(data_filepath, 'r') as fp:
        for line in fp:
            # ['am¥xc4¥x81bitis', 'indicative', 'active', 'future', 'pl', '2']
            surface_utf8, mood, voice, tense, number, person = line.rstrip().split('¥t')

            surface = surface_utf8.decode('utf-8')
            surface_v = vectorize_surface(surface)

            mood_v = vectorize_mood(mood) # [indicative, ...]
            voice_v = vectorize_voice(voice) # [active, passive]
            tense_v = vectorize_tense(tense) # [present, imperfect, future, perfect, past-perfect, future-perfect]
            number_v = vectorize_number(number) # [sg, pl]
            person_v = vectorize_person(person) # [1, 2, 3]

            x = flatten_list(surface_v)
            y = mood_v + voice_v + tense_v + number_v + person_v

            xs.append(x)
            ys.append(y)

    return np.array(xs), np.array(ys)


#if __name__ == '__main__':
#    X, Y = load_data(sys.argv[1])
#    print "X:", X
#    print "Y:", Y

nn.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
import numpy as np
import neurolab as nl
import sys

import latin_loader

if __name__ == '__main__':
    X, Y = latin_loader.load_data("latin.dat")

    w = X.shape[1]
    h = Y.shape[1]

    input_range = [[0, 1]] * w

    net = nl.net.newff(input_range, [25, h])
    err = net.train(X, Y, epochs=10, show=15) #, show=15)
    print err