[PYTHON] Apprenez Zundokokiyoshi en utilisant LSTM

Aperçu

J'ai essayé d'apprendre Zundokokiyoshi en utilisant LSTM. Il est implémenté à l'aide de Chainer. J'ai posté le mauvais code il y a quelque temps, mais je vais le réparer et le republier.

Le post suivant est détaillé pour l'explication de LSTM.

Comprendre le LSTM-avec les tendances récentes

modèle

Construisez un modèle comme celui ci-dessous

• Acceptez "Zun" ou "Doko" comme entrée
• La sortie est None ou "\ Kiyo Shi! /"
• Apprenez à sortir "\ ki yo shi! /" Si vous entrez "bouse", "bouse", "bouse", "bouse" et "doko" dans cet ordre, et sinon, affichez Aucun.

code

zundoko.py

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

import numpy as np
import chainer
from chainer import Variable, optimizers, functions as F, links as L

np.random.seed()
zun = 0
doko = 1
input_num = 2
input_words = ['Bouse', 'Doco']
none = 0
kiyoshi = 1
output_num = 2
output_words = [None, '\ Ki yo shi! /']
hidden_num = 8
update_iteration = 20

class Zundoko(chainer.Chain):
    def __init__(self):
        super(Zundoko, self).__init__(
            word=L.EmbedID(input_num, hidden_num),
            lstm=L.LSTM(hidden_num, hidden_num),
            linear=L.Linear(hidden_num, hidden_num),
            out=L.Linear(hidden_num, output_num),
        )

    def __call__(self, x, train=True):
        h1 = self.word(x)
        h2 = self.lstm(h1)
        h3 = F.relu(self.linear(h2))
        return self.out(h3)

    def reset_state(self):
        self.lstm.reset_state()

kiyoshi_list = [zun, zun, zun, zun, doko]
kiyoshi_pattern = 0
kiyoshi_mask = (1 << len(kiyoshi_list)) - 1
for token in kiyoshi_list:
    kiyoshi_pattern = (kiyoshi_pattern << 1) | token

zundoko = Zundoko()
for param in zundoko.params():
    data = param.data
    data[:] = np.random.uniform(-1, 1, data.shape)
optimizer = optimizers.Adam(alpha=0.01)
optimizer.setup(zundoko)

def forward(train=True):
    loss = 0
    acc = 0
    if train:
        batch_size = 20
    else:
        batch_size = 1
    recent_pattern = np.zeros((batch_size,), dtype=np.int32)
    zundoko.reset_state()
    for i in range(200):
        x = np.random.randint(0, input_num, batch_size).astype(np.int32)
        y_var = zundoko(Variable(x, volatile=not train), train=train)
        recent_pattern = ((recent_pattern << 1) | x) & kiyoshi_mask
        if i < len(kiyoshi_list):
            t = np.full((batch_size,), none, dtype=np.int32)
        else:
            t = np.where(recent_pattern == kiyoshi_pattern, kiyoshi, none).astype(np.int32)
        loss += F.softmax_cross_entropy(y_var, Variable(t, volatile=not train))
        acc += float(F.accuracy(y_var, Variable(t, volatile=not train)).data)
        if not train:
            print input_words[x[0]]
            y = np.argmax(y_var.data[0])
            if output_words[y] != None:
                print output_words[y]
                break
        if train and (i + 1) % update_iteration == 0:
            optimizer.zero_grads()
            loss.backward()
            loss.unchain_backward()
            optimizer.update()
            print 'train loss: {} accuracy: {}'.format(loss.data, acc / update_iteration)
            loss = 0
            acc = 0

for iteration in range(20):
    forward()

forward(train=False)

Exemple de sortie

train loss: 18.4753189087 accuracy: 0.020000000298
train loss: 16.216506958 accuracy: 0.0325000006706
train loss: 15.0742883682 accuracy: 0.0350000008941
train loss: 13.9205350876 accuracy: 0.385000001639
train loss: 12.5977449417 accuracy: 0.96249999404
(Omission)
train loss: 0.00433994689956 accuracy: 1.0
train loss: 0.00596862798557 accuracy: 1.0
train loss: 0.0027643663343 accuracy: 1.0
train loss: 0.011038181372 accuracy: 1.0
train loss: 0.00512072304264 accuracy: 1.0
Bouse
Bouse
Bouse
Doco
Doco
Doco
Bouse
Bouse
Bouse
Doco
Doco
Doco
Doco
Doco
Doco
Bouse
Doco
Doco
Bouse
Doco
Doco
Bouse
Doco
Bouse
Bouse
Bouse
Bouse
Bouse
Bouse
Bouse
Bouse
Doco
\ Ki yo shi! /

Un peu accro

Au début, j'utilisais le décrochage, mais l'apprentissage ne s'est pas bien déroulé et le résultat était presque nul.