def test_format_exceptions(self):
estream = RandomFixedSizeCrop(self.example_stream, (5, 4),
which_sources=('source2', ))
bstream = RandomFixedSizeCrop(self.batch_stream, (5, 4),
which_sources=('source2', ))
assert_raises(ValueError, estream.transform_source_example,
numpy.empty((5, 6)), 'source2')
assert_raises(ValueError, bstream.transform_source_batch,
[numpy.empty((7, 6))], 'source2')
assert_raises(ValueError, bstream.transform_source_batch,
[numpy.empty((8, 6))], 'source2')
def test_window_too_big_exceptions(self):
stream = RandomFixedSizeCrop(self.example_stream, (5, 4),
which_sources=('source2', ))
assert_raises(ValueError, stream.transform_source_example,
numpy.empty((3, 4, 2)), 'source2')
bstream = RandomFixedSizeCrop(self.batch_stream, (5, 4),
which_sources=('source1', ))
assert_raises(ValueError, bstream.transform_source_batch,
numpy.empty((5, 3, 4, 2)), 'source1')
def test_ndarray_batch_source(self):
# Make sure that with enough epochs we sample everything.
stream = RandomFixedSizeCrop(self.batch_stream, (5, 4),
which_sources=('source1', ))
seen_indices = numpy.array([], dtype='uint8')
for i in range(30):
for batch in stream.get_epoch_iterator():
assert batch[0].shape[1:] == (3, 5, 4)
assert batch[0].shape[0] in (1, 2)
seen_indices = numpy.union1d(seen_indices, batch[0].flatten())
if 3 * 7 * 5 == len(seen_indices):
break
else:
assert False
def test_ndarray_batch_source(self):
# Make sure that with enough epochs we sample everything.
stream = RandomFixedSizeCrop(self.batch_stream, (5, 4),
which_sources=('source1',))
seen_indices = numpy.array([], dtype='uint8')
for i in range(30):
for batch in stream.get_epoch_iterator():
assert batch[0].shape[1:] == (3, 5, 4)
assert batch[0].shape[0] in (1, 2)
seen_indices = numpy.union1d(seen_indices, batch[0].flatten())
if 3 * 7 * 5 == len(seen_indices):
break
else:
assert False
def test_list_batch_source(self):
# Make sure that with enough epochs we sample everything.
stream = RandomFixedSizeCrop(self.batch_stream, (5, 4),
which_sources=('source2', ))
seen_indices = numpy.array([], dtype='uint8')
for i in range(30):
for batch in stream.get_epoch_iterator():
for example in batch[1]:
assert example.shape == (2, 5, 4)
seen_indices = numpy.union1d(seen_indices,
example.flatten())
assert len(batch[1]) in (1, 2)
if self.source2_biggest == len(seen_indices):
break
else:
assert False
def test_list_batch_source(self):
# Make sure that with enough epochs we sample everything.
stream = RandomFixedSizeCrop(self.batch_stream, (5, 4),
which_sources=('source2',))
seen_indices = numpy.array([], dtype='uint8')
for i in range(30):
for batch in stream.get_epoch_iterator():
for example in batch[1]:
assert example.shape == (2, 5, 4)
seen_indices = numpy.union1d(seen_indices,
example.flatten())
assert len(batch[1]) in (1, 2)
if self.source2_biggest == len(seen_indices):
break
else:
assert False
train = DogsVsCats(('train', ), subset=slice(0, 20000))
# We now create a "stream" over the dataset which will return shuffled batches
# of size 128. Using the DataStream.default_stream constructor will turn our
# 8-bit images into floating-point decimals in [0, 1].
stream = DataStream.default_stream(train,
iteration_scheme=ShuffledScheme(
train.num_examples, 128))
# Enlarge images that are too small
downnscale_stream = MinimumImageDimensions(stream, (64, 64),
which_sources=('image_features', ))
# Our images are of different sizes, so we'll use a Fuel transformer
# to take random crops of size (32 x 32) from each image
cropped_stream = RandomFixedSizeCrop(downnscale_stream, (32, 32),
which_sources=('image_features', ))
# We'll use a simple MLP, so we need to flatten the images
# from (channel, width, height) to simply (features,)
flattened_stream = Flatten(cropped_stream, which_sources=('image_features', ))
# Create the Theano MLP
import theano
from theano import tensor
import numpy
X = tensor.matrix('image_features')
T = tensor.lmatrix('targets')
W = theano.shared(numpy.random.uniform(low=-0.01, high=0.01, size=(3072, 500)),
'W')
# We now create a "stream" over the dataset which will return shuffled batches
# of size 128. Using the DataStream.default_stream constructor will turn our
# 8-bit images into floating-point decimals in [0, 1].
stream = DataStream.default_stream(train,
iteration_scheme=SequentialScheme(
train.num_examples, batch_size))
# upscaled_stream = MinimumImageDimensions(stream, (100, 100), which_sources=('image_features',))
downscaled_stream = DownscaleMinDimension(stream,
100,
which_sources=('image_features', ))
# Our images are of different sizes, so we'll use a Fuel transformer
# to take random crops of size (32 x 32) from each image
cropped_stream = RandomFixedSizeCrop(downscaled_stream, (100, 100),
which_sources=('image_features', ))
rotated_stream = Random2DRotation(cropped_stream,
math.pi / 6,
which_sources=('image_features', ))
flipped_stream = RandomHorizontalFlip(rotated_stream,
which_sources=('image_features', ))
# We'll use a simple MLP, so we need to flatten the images
# from (channel, width, height) to simply (features,)
float_stream = ScaleAndShift(flipped_stream,
1. / 255,
0,
which_sources=('image_features', ))
float32_stream = Cast(float_stream,
numpy.float32,
