xor.py

# -*- coding: utf-8 -*-
"""
Chainerを用いた多層パーセプトロンの実装
多分に参考にしました: https://github.com/ichiroex/xor-mlp-chainer
"""
import sys
import numpy as np
import chainer
from chainer import cuda, Function, gradient_check, Variable, optimizers, serializers, utils
from chainer import Link, Chain, ChainList
import chainer.functions as F
import chainer.links as L

source = [[0, 0], [1, 0], [0, 1], [1, 1]]
target = [[0], [1], [1], [0]]
dataset = {}
dataset['source'] = np.array(source, dtype=np.float32)
dataset['target'] = np.array(target, dtype=np.float32)

class MyMLP(Chain):
    def __init__(self):
        super(MyMLP, self).__init__(
            l1 = L.Linear(2,2),
            l2 = L.Linear(2,1),
            )

    def forward(self, x, t):
        h1 = F.sigmoid(self.l1(x))
        return F.sigmoid(self.l2(h1))

if __name__ == "__main__":
    model = MyMLP()
    optimizer = optimizers.Adam()
    optimizer.setup(model)
    # Learning loop
    loss_val = 100
    epoch = 0
    while loss_val > 1e-5:
        # training
        x = chainer.Variable(np.asarray(dataset['source'])) #source
        t = chainer.Variable(np.asarray(dataset['target'])) #target
        model.zerograds()       # 勾配をゼロ初期化
        y = model.forward(x,t)
        loss = F.mean_squared_error(y, t)
        loss.backward()              # 誤差逆伝播
        optimizer.update()           # 最適化 
        # 途中結果を表示
        if epoch % 1000 == 0:
            #誤差と正解率を計算
            loss_val = loss.data
            print 'epoch:', epoch
            print 'x:\n', x.data
            print 't:\n', t.data
            print 'y:\n', y.data
            print('train mean loss={}'.format(loss_val)) # 訓練誤差, 正解率
            print ' - - - - - - - - - '
        # n_epoch以上になると終了
        if epoch >= 100000:
            break
        epoch += 1