Different behaviour in gpu and cpu.

simpler.py

import theano
import numpy

from collections import OrderedDict
from fuel.datasets import IndexableDataset
from fuel.streams import DataStream
from fuel.schemes import SequentialScheme

from theano import tensor

from blocks.monitoring.evaluators import DatasetEvaluator
from blocks.monitoring.aggregation import Mean

num_examples = 5
num_batches = 3
batch_size = 2
features = numpy.array([[3, 3],
                       [2, 9],
                       [2, 4],
                       [5, 1],
                       [3, 3]], dtype=theano.config.floatX)

print -features.mean(axis = 0)
#[-3. -4.]

dataset = IndexableDataset(OrderedDict([('features', features)]))

x = tensor.matrix('features')
y = -x
y.name = 'y'

y.tag.aggregation_scheme = Mean(y.sum(axis=0), y.shape[0])

data_stream_2 = DataStream(dataset,
                           iteration_scheme=SequentialScheme(num_examples, 2))

print DatasetEvaluator([y]).evaluate(data_stream_2)['y']