gz_nn_minibatch.py

from pybrain.datasets import ClassificationDataSet
from pybrain.tools.shortcuts import buildNetwork
from pybrain.supervised.trainers import BackpropTrainer
from pybrain.structure.modules import SoftmaxLayer
from sklearn import datasets
import random

# load the data from sklearn
iris = datasets.load_iris()
X = iris['data']
y = iris['target']

# set up the dataset. 4 input features, 3 output classes
all_inds = range(X.shape[0])

# build the network
fnn = buildNetwork(4, 10, 3, outclass=SoftmaxLayer, bias=True)
trainer = BackpropTrainer(fnn, momentum=0.1, verbose=True, weightdecay=0.01, learningrate=0.01)

# repeat the batch training several times
for i in xrange(200):
    # get a random order for the training examples for batch gradient descent
    random.shuffle(all_inds)
    # split the indexes into lists with the indices for each batch
    batch_inds = [all_inds[i:i+10] for i in xrange(0, len(all_inds), 10)]

    # train on each batch
    for inds in batch_inds:
        # rebuild the dataset
        ds = ClassificationDataSet(4, nb_classes=3)
        for x_i, y_i in zip(X[inds, :], y[inds]):
            ds.appendLinked(x_i, y_i)
        ds._convertToOneOfMany()
        # train on the current batch
        trainer.trainOnDataset(ds)

# make a dataset with all the iris data
ds_all = ClassificationDataSet(4, nb_classes=3)
for x_i, y_i in zip(X, y):
    ds_all.appendLinked(x_i, y_i)
ds_all._convertToOneOfMany()

# test the result
# Note that we are testing on our training data, which is bad practice,
# but it does demonstrate the network is trained
print sum(fnn.activateOnDataset(ds_all).argmax(axis=1) == y)/float(len(y))