test_chainer.py

#!/usr/bin/env python
import numpy as np
import six

import chainer
from chainer import computational_graph as c
from chainer import cuda
import chainer.functions as F
from chainer import optimizers

inputlayersize = 400
hiddenlayersize = 25
numlabel = 10

# Data files are available from the following URLs.
# * https://dl.dropboxusercontent.com/u/123796/gist/b553004a7ca76310e11c/data/ex4data1_X.csv.gz
# * https://dl.dropboxusercontent.com/u/123796/gist/b553004a7ca76310e11c/data/ex4data1_y.csv
# 
# They are converted from ex4data1.mat using Octave as below.
# 
# > load('ex4data1.mat');
# > csvwrite("ex4data1_X.csv", X);
# > csvwrite("ex4data1_y.csv", y);
x = np.loadtxt("ex4data1_X.csv", delimiter=",", dtype=np.float32)
y = np.loadtxt("ex4data1_y.csv", delimiter=",", dtype=np.int_)
y = np.vectorize(lambda x: 0 if x == 10 else x, otypes = [np.int_])(y)

# Prepare multi-layer perceptron model
model = chainer.FunctionSet(l1=F.Linear(inputlayersize, hiddenlayersize),
                            l2=F.Linear(hiddenlayersize, numlabel))

# Neural net architecture

def forward(x):
    x = chainer.Variable(x)
    return F.sigmoid(model.l2(F.sigmoid(model.l1(x))))

def cost(x, y):
    o = forward(x)
    y2 = np.zeros([len(y), numlabel], np.float32)
    for i in xrange(y.shape[0]):
        y2[i, y[i]] = 1
    c = F.sum(F.sum(- y2 * F.log(o) - (1 - y2) * F.log(1 - o))) / len(y)
    return c

def predict(x):
    return forward(x).data.argmax(axis=1)

# Setup optimizer
optimizer = optimizers.SGD()
optimizer.setup(model.collect_parameters())

for i in xrange(1000):
    o = cost(x, y)
    print o.data
    optimizer.zero_grads()
    o.backward()
    optimizer.update()

p = predict(x)
print("precision (training data): %f" % (sum(1.0 for i in xrange(y.shape[0]) if p[i] == y[i]) / y.shape[0]))