def test_ngram_stream_error_on_multiple_sources():
# Check that NGram accepts only data streams with one source
sentences = [list(numpy.random.randint(10, size=sentence_length))
for sentence_length in [3, 5, 7]]
stream = DataStream(IterableDataset(sentences))
stream.sources = ('1', '2')
assert_raises(ValueError, NGrams, 4, stream)
def load_datastream(train_batch_size=100):
from fuel.datasets.mnist import MNIST
from fuel.transformers import ScaleAndShift, Cast, Flatten, Mapping
from fuel.streams import DataStream
from fuel.schemes import SequentialScheme, ShuffledScheme
MNIST.default_transformers = (
(ScaleAndShift, [2.0 / 255.0, -1], {'which_sources': 'features'}),
(Cast, [np.float32], {'which_sources': 'features'}),
)
mnist_train = MNIST(('train',), subset=slice(None, 50000))
mnist_train_stream = DataStream.default_stream(
mnist_train,
iteration_scheme=ShuffledScheme(mnist_train.num_examples, train_batch_size)
)
mnist_validation = MNIST(('train',), subset=slice(50000, None))
mnist_validation_stream = DataStream.default_stream(
mnist_validation,
iteration_scheme=SequentialScheme(mnist_validation.num_examples, 250)
)
mnist_test = MNIST(('test',))
mnist_test_stream = DataStream.default_stream(
mnist_test,
iteration_scheme=SequentialScheme(mnist_test.num_examples, 250)
)
return {
'train': mnist_train_stream,
'validation': mnist_validation_stream,
'test': mnist_test_stream
}
def get_streams(num_train_examples, batch_size, use_test=True):
dataset = MNIST(("train",))
all_ind = numpy.arange(dataset.num_examples)
rng = numpy.random.RandomState(seed=1)
rng.shuffle(all_ind)
indices_train = all_ind[:num_train_examples]
indices_valid = all_ind[num_train_examples:]
tarin_stream = Flatten(
DataStream.default_stream(dataset, iteration_scheme=ShuffledScheme(indices_train, batch_size))
)
valid_stream = None
if len(indices_valid) != 0:
valid_stream = Flatten(
DataStream.default_stream(dataset, iteration_scheme=ShuffledScheme(indices_valid, batch_size))
)
test_stream = None
if use_test:
dataset = MNIST(("test",))
ind = numpy.arange(dataset.num_examples)
rng = numpy.random.RandomState(seed=1)
rng.shuffle(all_ind)
test_stream = Flatten(DataStream.default_stream(dataset, iteration_scheme=ShuffledScheme(ind, batch_size)))
return tarin_stream, valid_stream, test_stream
def test_in_memory():
skip_if_not_available(datasets=['mnist.hdf5'])
# Load MNIST and get two batches
mnist = MNIST(('train',), load_in_memory=True)
data_stream = DataStream(mnist, iteration_scheme=SequentialScheme(
examples=mnist.num_examples, batch_size=256))
epoch = data_stream.get_epoch_iterator()
for i, (features, targets) in enumerate(epoch):
if i == 1:
break
handle = mnist.open()
known_features, _ = mnist.get_data(handle, slice(256, 512))
mnist.close(handle)
assert numpy.all(features == known_features)
# Pickle the epoch and make sure that the data wasn't dumped
with tempfile.NamedTemporaryFile(delete=False) as f:
filename = f.name
cPickle.dump(epoch, f)
assert os.path.getsize(filename) < 1024 * 1024 # Less than 1MB
# Reload the epoch and make sure that the state was maintained
del epoch
with open(filename, 'rb') as f:
epoch = cPickle.load(f)
features, targets = next(epoch)
handle = mnist.open()
known_features, _ = mnist.get_data(handle, slice(512, 768))
mnist.close(handle)
assert numpy.all(features == known_features)
def maxout_vae_mnist_test(path_vae_mnist):
# load vae model on mnist
vae_mnist = load(path_vae_mnist)
maxout = Maxout()
x = T.matrix('features')
y = T.imatrix('targets')
batch_size = 128
z, _ = vae_mnist.sampler.sample(vae_mnist.encoder_mlp.apply(x))
predict = maxout.apply(z)
cost = Softmax().categorical_cross_entropy(y.flatten(), predict)
y_hat = Softmax().apply(predict)
cost.name = 'cost'
cg = ComputationGraph(cost)
temp = cg.parameters
for t, i in zip(temp, range(len(temp))):
t.name = t.name+str(i)+"maxout"
error_brick = MisclassificationRate()
error_rate = error_brick.apply(y, y_hat)
# training
step_rule = RMSProp(0.01, 0.9)
#step_rule = Momentum(0.2, 0.9)
train_set = MNIST('train')
test_set = MNIST("test")
data_stream_train = Flatten(DataStream.default_stream(
train_set, iteration_scheme=SequentialScheme(train_set.num_examples, batch_size)))
data_stream_test =Flatten(DataStream.default_stream(
test_set, iteration_scheme=SequentialScheme(test_set.num_examples, batch_size)))
algorithm = GradientDescent(cost=cost, params=cg.parameters,
step_rule=step_rule)
monitor_train = TrainingDataMonitoring(
variables=[cost], data_stream=data_stream_train, prefix="train")
monitor_valid = DataStreamMonitoring(
variables=[cost, error_rate], data_stream=data_stream_test, prefix="test")
extensions = [ monitor_train,
monitor_valid,
FinishAfter(after_n_epochs=50),
Printing(every_n_epochs=1)
]
main_loop = MainLoop(data_stream=data_stream_train,
algorithm=algorithm, model = Model(cost),
extensions=extensions)
main_loop.run()
# save here
from blocks.serialization import dump
with closing(open('../data_mnist/maxout', 'w')) as f:
dump(maxout, f)
def test_unpack_transformer():
data = range(10)
stream = DataStream(IterableDataset(data))
stream = Batch(stream, iteration_scheme=ConstantScheme(2))
stream = Unpack(stream)
epoch = stream.get_epoch_iterator()
for i, v in enumerate(epoch):
assert numpy.shape(v)[0] == 1
assert v[0] == i
def prepare_cifar10():
class Dataset:
pass
result = Dataset()
CIFAR10.default_transformers = (
(ScaleAndShift, [2.0 / 255.0, -1], {'which_sources': 'features'}),
(Cast, [np.float32], {'which_sources': 'features'}))
mean = cifar10_mean()
def patch_get_epoch_iterator(stream):
def get_epoch_iterator(self):
for X, Y in self._get_epoch_iterator():
# 0 degrees
X -= mean[numpy.newaxis,:,:,:]
yield augument(X, 25), Y
stream._get_epoch_iterator = stream.get_epoch_iterator
stream.get_epoch_iterator = types.MethodType(get_epoch_iterator, stream)
def patch_get_epoch_iterator_test(stream):
def get_epoch_iterator(self):
for X, Y in self._get_epoch_iterator():
# 0 degrees
X -= mean[numpy.newaxis,:,:,:]
yield X, Y
stream._get_epoch_iterator = stream.get_epoch_iterator
stream.get_epoch_iterator = types.MethodType(get_epoch_iterator, stream)
result.train = train = CIFAR10(("train",), subset = slice(None, 40000))
result.train_stream = DataStream.default_stream(
result.train,
iteration_scheme = ShuffledScheme(result.train.num_examples, 25))
patch_get_epoch_iterator(result.train_stream)
result.validation = CIFAR10(("train",), subset=slice(40000, None))
result.validation_stream = DataStream.default_stream(
result.validation,
iteration_scheme = SequentialScheme(result.validation.num_examples, 100))
patch_get_epoch_iterator(result.validation_stream)
result.test = CIFAR10(("test",))
result.test_stream = DataStream.default_stream(
result.test,
iteration_scheme = SequentialScheme(result.test.num_examples, 100))
patch_get_epoch_iterator_test(result.test_stream)
return result
def test_get_data_inmemory(self):
ds = ImagesFromFile('{}/*.jpeg'.format(self.ds_path),
load_in_memory=True)
stream = DataStream(ds, iteration_scheme=ShuffledScheme(
ds.num_examples, batch_size=10))
for imgs, file_paths in stream.get_epoch_iterator():
assert len(imgs) == 5
assert all([img.shape == (512, 512, 3) for img in imgs])
assert all([self.ds_path in fp for fp in file_paths])
assert len(set(file_paths)) == 5
def test_batch_iteration_scheme_with_lists(self):
"""Batch schemes should work with more than ndarrays."""
data = IndexableDataset(OrderedDict([('foo', list(range(50))),
('bar', list(range(1, 51)))]))
stream = DataStream(data,
iteration_scheme=ShuffledScheme(data.num_examples,
5))
returned = [sum(batches, []) for batches in
zip(*list(stream.get_epoch_iterator()))]
assert set(returned[0]) == set(range(50))
assert set(returned[1]) == set(range(1, 51))
def __init__(self,dataset,batch_size=128,shuffle=False):
self.dataset = dataset
if shuffle:
self.datastream = DataStream(self.dataset,
iteration_scheme=ShuffledScheme(
examples=dataset.num_examples,
batch_size=batch_size))
else:
self.datastream = DataStream(self.dataset,
iteration_scheme=SequentialScheme(
examples=dataset.num_examples,
batch_size=batch_size))
def main(save_to, num_epochs, bokeh=False):
mlp = MLP([Tanh(), Softmax()], [784, 100, 10],
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0))
mlp.initialize()
x = tensor.matrix('features')
y = tensor.lmatrix('targets')
probs = mlp.apply(tensor.flatten(x, outdim=2))
cost = CategoricalCrossEntropy().apply(y.flatten(), probs)
error_rate = MisclassificationRate().apply(y.flatten(), probs)
cg = ComputationGraph([cost])
W1, W2 = VariableFilter(roles=[WEIGHT])(cg.variables)
cost = cost + .00005 * (W1 ** 2).sum() + .00005 * (W2 ** 2).sum()
cost.name = 'final_cost'
mnist_train = MNIST("train")
mnist_test = MNIST("test")
algorithm = GradientDescent(
cost=cost, params=cg.parameters,
step_rule=Scale(learning_rate=0.1))
extensions = [Timing(),
FinishAfter(after_n_epochs=num_epochs),
DataStreamMonitoring(
[cost, error_rate],
Flatten(
DataStream.default_stream(
mnist_test,
iteration_scheme=SequentialScheme(
mnist_test.num_examples, 500)),
which_sources=('features',)),
prefix="test"),
TrainingDataMonitoring(
[cost, error_rate,
aggregation.mean(algorithm.total_gradient_norm)],
prefix="train",
after_epoch=True),
Checkpoint(save_to),
Printing()]
main_loop = MainLoop(
algorithm,
Flatten(
DataStream.default_stream(
mnist_train,
iteration_scheme=SequentialScheme(
mnist_train.num_examples, 50)),
which_sources=('features',)),
model=Model(cost),
extensions=extensions)
main_loop.run()
def test_dataset():
data = [1, 2, 3]
# The default stream requests an example at a time
stream = DataStream(IterableDataset(data))
epoch = stream.get_epoch_iterator()
assert list(epoch) == list(zip(data))
# Check if iterating over multiple epochs works
for i, epoch in zip(range(2), stream.iterate_epochs()):
assert list(epoch) == list(zip(data))
# Check whether the returning as a dictionary of sources works
assert next(stream.get_epoch_iterator(as_dict=True)) == {"data": 1}
def batch_iterator(self, dataset, batchsize, shuffle=False):
if isinstance(dataset, Dataset):
if shuffle:
train_scheme = ShuffledScheme(examples=dataset.num_examples, batch_size=batchsize)
else:
train_scheme = SequentialScheme(examples=dataset.num_examples, batch_size=batchsize)
stream = DataStream(dataset=dataset, iteration_scheme=train_scheme)
if self.fuel_stream_xform_fn is not None:
stream = self.fuel_stream_xform_fn(stream)
return stream.get_epoch_iterator()
elif _is_sequence_of_arrays(dataset):
return iterate_minibatches(dataset, batchsize, shuffle=shuffle)
else:
raise TypeError('dataset should be a fuel Dataset instance or a list of arrays')
def prep_dataset(dataset):
dataset = dataset[:(len(dataset) - (len(dataset) % (seq_len * batch_size)))]
dataset = dataset.reshape(batch_size, -1, seq_len).transpose((1, 0, 2))
stream = DataStream(IndexableDataset(indexables=OrderedDict([
('data', dataset)])),
iteration_scheme=SequentialExampleScheme(dataset.shape[0]))
stream = Transpose(stream, [(1, 0)])
stream = SampleDropsNPWord(
stream, z_prob_states, z_prob_cells, drop_prob_igates,
layer_size, num_layers, False, stoch_depth, share_mask,
gaussian_drop, alphabetsize)
stream.sources = ('data',) * 3 + stream.sources + ('zoneouts_states', 'zoneouts_cells', 'zoneouts_igates')
return (stream,)
def get_stream(batch_size, input_size, test=False):
from fuel.datasets.dogs_vs_cats import DogsVsCats
from fuel.streams import DataStream
from fuel.schemes import ShuffledScheme, SequentialScheme ,SequentialExampleScheme
from fuel.transformers.image import RandomFixedSizeCrop
from fuel.transformers import Flatten #, ForceFloatX
from ScikitResize import ScikitResize
from fuel.transformers import Cast
# Load the training set
if test :
train = DogsVsCats(('train',),subset=slice(0, 30))
valid = DogsVsCats(('train',),subset=slice(19980, 20000))
test = DogsVsCats(('test',),subset=slice(0,4))
else :
train = DogsVsCats(('train',),subset=slice(0,22000))
valid = DogsVsCats(('train',),subset=slice(22000, 25000))
test = DogsVsCats(('test',))
#Generating stream
train_stream = DataStream.default_stream(
train,
iteration_scheme=ShuffledScheme(train.num_examples, batch_size)
)
valid_stream = DataStream.default_stream(
valid,
iteration_scheme=ShuffledScheme(valid.num_examples, batch_size)
)
test_stream = DataStream.default_stream(
test,
iteration_scheme=SequentialScheme(test.num_examples, 1)
# iteration_scheme=SequentialExampleScheme(test.num_examples)
)
#Reshaping procedure
#Apply crop and resize to desired square shape
train_stream = ScikitResize(train_stream, input_size, which_sources=('image_features',))
valid_stream = ScikitResize(valid_stream, input_size, which_sources=('image_features',))
test_stream = ScikitResize(test_stream, input_size, which_sources=('image_features',))
#ForceFloatX, to spare you from possible bugs
#train_stream = ForceFloatX(train_stream)
#valid_stream = ForceFloatX(valid_stream)
#test_stream = ForceFloatX(test_stream)
#Cast instead of forcefloatX
train_stream = Cast(train_stream, dtype='float32',which_sources=('image_features',))
valid_stream = Cast(valid_stream, dtype='float32',which_sources=('image_features',))
test_stream = Cast(test_stream, dtype='float32',which_sources=('image_features',))
return train_stream, valid_stream, test_stream
def test_h5py_dataset_datastream_pickles():
try:
h5file = h5py.File(name='tmp.hdf5', mode="w")
features = h5file.create_dataset('features', (10, 5), dtype='float32')
features[...] = numpy.arange(50, dtype='float32').reshape((10, 5))
split_dict = {'train': {'features': (0, 10)}}
h5file.attrs['split'] = H5PYDataset.create_split_array(split_dict)
h5file.flush()
h5file.close()
dataset = H5PYDataset(path='tmp.hdf5', which_set='train')
stream = DataStream(dataset)
pickle.loads(pickle.dumps(stream))
finally:
stream.close()
if os.path.exists('tmp.hdf5'):
os.remove('tmp.hdf5')
def monk_music_stream (which_sets = ('train',),batch_size = 64,
seq_size=128, frame_size=160, num_examples= None,
which_sources = ('features',)):
"""
This function generates the stream for the monk_music dataset.
It doesn't compute incremental windows and instead simply separates the
dataset into sequences
"""
dataset = MonkMusic(which_sets = which_sets, filename = "dataset.hdf5",
load_in_memory=True)
large_batch_size = batch_size * frame_size * seq_size
if not num_examples:
num_examples = large_batch_size*(dataset.num_examples/large_batch_size)
# If there are memory problems revert to SequentialScheme
data_stream = DataStream.default_stream(
dataset, iteration_scheme=SequentialScheme(
num_examples,
large_batch_size))
data_stream = ScaleAndShift(data_stream,
scale = 1./data_stats["std"],
shift = -data_stats["mean"]/data_stats["std"])
data_stream = Mapping(data_stream,
lambda data: _get_subsequences(data,batch_size,seq_size,frame_size))
data_stream = ForceFloatX(data_stream)
return data_stream
def get_mnist_video_streams(batch_size):
train_dataset = ClutteredMNISTVideo(which_sets=["train"])
valid_dataset = ClutteredMNISTVideo(which_sets=["valid"])
train_ind = numpy.arange(train_dataset.num_examples)
valid_ind = numpy.arange(valid_dataset.num_examples)
rng = numpy.random.RandomState(seed=1)
rng.shuffle(train_ind)
rng.shuffle(valid_ind)
train_datastream = DataStream.default_stream(train_dataset, iteration_scheme=ShuffledScheme(train_ind, batch_size))
train_datastream = PreprocessTransformer(train_datastream)
valid_datastream = DataStream.default_stream(valid_dataset, iteration_scheme=ShuffledScheme(valid_ind, batch_size))
valid_datastream = PreprocessTransformer(valid_datastream)
return train_datastream, valid_datastream
def create_dataset(dataset):
if trainning:
scheme = ShuffledScheme(dataset.num_examples, 32)
else:
scheme = SequentialScheme(dataset.num_examples, 32)
stream = DataStream.default_stream(dataset, iteration_scheme=scheme)
return ResizeTransformer(stream, image_size)
def test_cifar100():
train = CIFAR100('train', load_in_memory=False)
assert train.num_examples == 50000
handle = train.open()
coarse_labels, features, fine_labels = train.get_data(handle,
slice(49990, 50000))
assert features.shape == (10, 3, 32, 32)
assert coarse_labels.shape == (10, 1)
assert fine_labels.shape == (10, 1)
train.close(handle)
test = CIFAR100('test', load_in_memory=False)
handle = test.open()
coarse_labels, features, fine_labels = test.get_data(handle,
slice(0, 10))
assert features.shape == (10, 3, 32, 32)
assert coarse_labels.shape == (10, 1)
assert fine_labels.shape == (10, 1)
assert features.dtype == numpy.uint8
assert coarse_labels.dtype == numpy.uint8
assert fine_labels.dtype == numpy.uint8
test.close(handle)
stream = DataStream.default_stream(
test, iteration_scheme=SequentialScheme(10, 10))
data = next(stream.get_epoch_iterator())[1]
assert data.min() >= 0.0 and data.max() <= 1.0
assert data.dtype == config.floatX
assert_raises(ValueError, CIFAR100, 'valid')
def DStream(datatype, config):
if datatype=='train':
filename = config['train_file']
elif datatype == 'valid':
filename = config['valid_file']
elif datatype == 'test':
filename = config['test_file']
else:
logger.error('wrong datatype, train, valid, or test')
data = TextFile(files=[filename],
dictionary=pickle.load(open(config['train_dic'],'rb')),
unk_token=config['unk_token'],
level='word',
bos_token=config['bos_token'],
eos_token=config['eos_token'])
data_stream = DataStream.default_stream(data)
data_stream.sources = ('sentence',)
# organize data in batches and pad shorter sequences with zeros
batch_size = config['batch_size']
data_stream = Batch(data_stream, iteration_scheme=ConstantScheme(batch_size))
data_stream = Padding(data_stream)
return data_stream
def test_cifar10():
train = CIFAR10(('train',), load_in_memory=False)
assert train.num_examples == 50000
handle = train.open()
features, targets = train.get_data(handle, slice(49990, 50000))
assert features.shape == (10, 3, 32, 32)
assert targets.shape == (10, 1)
train.close(handle)
test = CIFAR10(('test',), load_in_memory=False)
handle = test.open()
features, targets = test.get_data(handle, slice(0, 10))
assert features.shape == (10, 3, 32, 32)
assert targets.shape == (10, 1)
assert features.dtype == numpy.uint8
assert targets.dtype == numpy.uint8
test.close(handle)
stream = DataStream.default_stream(
test, iteration_scheme=SequentialScheme(10, 10))
data = next(stream.get_epoch_iterator())[0]
assert data.min() >= 0.0 and data.max() <= 1.0
assert data.dtype == config.floatX
assert_raises(ValueError, CIFAR10, ('valid',))
dummy = CIFAR10(('train',), subset=slice(50000, 60000))
handle = dummy.open()
assert_raises(ValueError, dummy.get_data, handle, slice(0, 10000))
dummy.close(handle)
def open_stream(which_sets= ('train',), port=5557, num_examples = None):
dataset = Blizzard(which_sets = which_sets)
if num_examples == None:
num_examples = dataset.num_examples
data_stream = DataStream.default_stream(
dataset, iteration_scheme=SequentialScheme(
num_examples, batch_size))
data_stream = ScaleAndShift(data_stream, scale = 1/data_std,
shift = -data_mean/data_std)
data_stream = Mapping(data_stream, _downsample_and_upsample,
add_sources=('upsampled',))
data_stream = Mapping(data_stream, _equalize_size)
data_stream = Mapping(data_stream, _get_residual,
add_sources = ('residual',))
data_stream = FilterSources(data_stream,
sources = ('upsampled', 'residual',))
data_stream = Mapping(data_stream, _segment_axis)
data_stream = Mapping(data_stream, _transpose)
data_stream = ForceFloatX(data_stream)
start_server(data_stream, port=port)
def get_cmv_v1_streams(batch_size):
train_dataset = CMVv1(which_sets=["train"])
valid_dataset = CMVv1(which_sets=["valid"])
train_ind = numpy.arange(train_dataset.num_examples)
valid_ind = numpy.arange(valid_dataset.num_examples)
rng = numpy.random.RandomState(seed=1)
rng.shuffle(train_ind)
rng.shuffle(valid_ind)
train_datastream = DataStream.default_stream(train_dataset, iteration_scheme=ShuffledScheme(train_ind, batch_size))
train_datastream = Preprocessor_CMV_v1(train_datastream)
valid_datastream = DataStream.default_stream(valid_dataset, iteration_scheme=ShuffledScheme(valid_ind, batch_size))
valid_datastream = Preprocessor_CMV_v1(valid_datastream)
return train_datastream, valid_datastream
def fuel_data_to_list(fuel_data, shuffle):
if(shuffle):
scheme = ShuffledScheme(fuel_data.num_examples, fuel_data.num_examples)
else:
scheme = SequentialScheme(fuel_data.num_examples, fuel_data.num_examples)
fuel_data_stream = DataStream.default_stream(fuel_data, iteration_scheme=scheme)
return fuel_data_stream.get_epoch_iterator().next()
def create_svhn_data_streams(batch_size, monitoring_batch_size, rng=None):
train_set = SVHN(2, ('extra',), sources=('features',))
valid_set = SVHN(2, ('train',), sources=('features',))
main_loop_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(
train_set.num_examples, batch_size, rng=rng))
train_monitor_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(
5000, monitoring_batch_size, rng=rng))
valid_monitor_stream = DataStream.default_stream(
valid_set,
iteration_scheme=ShuffledScheme(
5000, monitoring_batch_size, rng=rng))
return main_loop_stream, train_monitor_stream, valid_monitor_stream
def create_tiny_imagenet_data_streams(batch_size, monitoring_batch_size,
rng=None):
train_set = TinyILSVRC2012(('train',), sources=('features',))
valid_set = TinyILSVRC2012(('valid',), sources=('features',))
main_loop_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(
train_set.num_examples, batch_size, rng=rng))
train_monitor_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(
4096, monitoring_batch_size, rng=rng))
valid_monitor_stream = DataStream.default_stream(
valid_set,
iteration_scheme=ShuffledScheme(
4096, monitoring_batch_size, rng=rng))
return main_loop_stream, train_monitor_stream, valid_monitor_stream
def set_datastream(data_path, batch_size):
dataset = H5PYDataset(file_or_path=data_path,
which_sets=('train',),
sources=('input_feature', 'target_feature'))
data_stream = DataStream.default_stream(dataset=dataset,
iteration_scheme=ShuffledScheme(batch_size=batch_size,
examples=dataset.num_examples))
return data_stream
def create_celeba_data_streams(batch_size, monitoring_batch_size,
sources=('features', ), rng=None):
train_set = CelebA('64', ('train',), sources=sources)
valid_set = CelebA('64', ('valid',), sources=sources)
main_loop_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(
train_set.num_examples, batch_size, rng=rng))
train_monitor_stream = DataStream.default_stream(
train_set,
iteration_scheme=ShuffledScheme(
5000, monitoring_batch_size, rng=rng))
valid_monitor_stream = DataStream.default_stream(
valid_set,
iteration_scheme=ShuffledScheme(
5000, monitoring_batch_size, rng=rng))
return main_loop_stream, train_monitor_stream, valid_monitor_stream
def create_gaussian_mixture_data_streams(batch_size,
monitoring_batch_size,
means=None,
variances=None,
priors=None,
rng=None,
num_examples=100000,
sources=('features', )):
train_set = GaussianMixture(num_examples=num_examples,
means=means,
variances=variances,
priors=priors,
rng=rng,
sources=sources)
valid_set = GaussianMixture(num_examples=num_examples,
means=means,
variances=variances,
priors=priors,
rng=rng,
sources=sources)
main_loop_stream = DataStream(train_set,
iteration_scheme=ShuffledScheme(
train_set.num_examples,
batch_size=batch_size,
rng=rng))
train_monitor_stream = DataStream(train_set,
iteration_scheme=ShuffledScheme(
5000, batch_size, rng=rng))
valid_monitor_stream = DataStream(valid_set,
iteration_scheme=ShuffledScheme(
5000, batch_size, rng=rng))
return main_loop_stream, train_monitor_stream, valid_monitor_stream
def get_sgnmt_shuffled_tr_stream(src_data,
trg_data,
src_vocab_size=30000,
trg_vocab_size=30000,
unk_id=1,
seq_len=50,
batch_size=80,
sort_k_batches=12,
**kwargs):
"""Prepares the shuffled training data stream. This is similar to
``get_sgnmt_tr_stream`` but uses ``ParallelTextFile`` in combination
with ``ShuffledExampleScheme`` to support reshuffling."""
# Build dummy vocabulary to make TextFile happy
src_vocab = add_special_ids({str(i): i for i in xrange(src_vocab_size)})
trg_vocab = add_special_ids({str(i): i for i in xrange(trg_vocab_size)})
parallel_dataset = ParallelTextFile(src_data, trg_data, src_vocab,
trg_vocab, None)
#iter_scheme = SequentialExampleScheme(parallel_dataset.num_examples)
iter_scheme = ShuffledExampleScheme(parallel_dataset.num_examples)
s = DataStream(parallel_dataset, iteration_scheme=iter_scheme)
# Filter sequences that are too long
s = Filter(s, predicate=stream._too_long(seq_len=seq_len))
# Replace out of vocabulary tokens with unk token
s = Mapping(
s,
stream._oov_to_unk(src_vocab_size=src_vocab_size,
trg_vocab_size=trg_vocab_size,
unk_id=utils.UNK_ID))
# Build a batched version of stream to read k batches ahead
s = Batch(s, iteration_scheme=ConstantScheme(batch_size * sort_k_batches))
# Sort all samples in the read-ahead batch
s = Mapping(s, SortMapping(stream._length))
# Convert it into a stream again
s = Unpack(s)
# Construct batches from the stream with specified batch size
s = Batch(s, iteration_scheme=ConstantScheme(batch_size))
# Pad sequences that are short
masked_stream = stream.PaddingWithEOS(s, [utils.EOS_ID, utils.EOS_ID])
return masked_stream
def timit_datastream(path, which_set, local_copy, pool_size, maximum_frames):
# load dataset
timit_dataset = Timit(which_set=which_set,
path=path,
local_copy=local_copy)
# get statistics
data_means, data_stds = timit_dataset.get_normalization_factors()
# set shuffle range
shuffle_rng = numpy.random.RandomState(123)
# set iterator scheme
iterator_scheme = SequentialShuffledScheme(
num_examples=timit_dataset.num_examples,
batch_size=pool_size,
rng=shuffle_rng)
# base data stream
base_stream = DataStream(dataset=timit_dataset,
iteration_scheme=iterator_scheme)
# reshape stream
reshape_stream = Reshape(data_source='features',
shape_source='features_shapes',
data_stream=base_stream)
# normalize data stream
normalize_stream = Normalize(data_stream=reshape_stream,
means=data_means,
stds=data_stds)
# sort data stream
sort_stream = Mapping(data_stream=normalize_stream,
mapping=SortMapping(key=lambda x: x[0].shape[0]))
# max frame stream
max_frame_stream = MaximumFrameCache(max_frames=maximum_frames,
data_stream=sort_stream,
rng=shuffle_rng)
# padding data stream
padded_stream = Padding(data_stream=max_frame_stream,
mask_sources=['features', 'phonemes'])
# floatX stream
data_stream = ForceFloatX(padded_stream)
return timit_dataset, data_stream
def create_data_generator(path, vocab_file, config):
ds = QADataset(path,
vocab_file,
config.n_entities,
need_sep_token=config.concat_ctx_and_question)
it = QAIterator(path, shuffle=config.shuffle_questions)
stream = DataStream(ds, iteration_scheme=it)
if config.concat_ctx_and_question:
stream = ConcatCtxAndQuestion(stream, config.concat_question_before,
ds.reverse_vocab['<SEP>'])
# Sort sets of multiple batches to make batches of similar sizes
stream = Batch(stream,
iteration_scheme=ConstantScheme(config.batch_size *
config.sort_batch_count))
comparison = _balanced_batch_helper(
stream.sources.index(
'question' if config.concat_ctx_and_question else 'context'))
stream = Mapping(stream, SortMapping(comparison))
stream = Unpack(stream)
stream = Batch(stream, iteration_scheme=ConstantScheme(config.batch_size))
stream = Padding(stream,
mask_sources=['context', 'question', 'candidates'],
mask_dtype='int32')
def gen():
if not config.concat_ctx_and_question:
for (seq_cont, seq_cont_mask, seq_quest, seq_quest_mask, tg,
candidates, candidates_mask) in stream.get_epoch_iterator():
seq_cont_mask = seq_cont_mask.astype('float32')
seq_quest_mask = seq_quest_mask.astype('float32')
candidates_mask = candidates_mask.astype('float32')
yield (seq_cont, seq_cont_mask, seq_quest, seq_quest_mask, tg,
candidates, candidates_mask)
else:
for (seq, seq_mask, tg, candidates, candidates_mask) \
in stream.get_epoch_iterator():
seq_mask = seq_mask.astype('float32')
candidates_mask = candidates_mask.astype('float32')
yield (seq, seq_mask, tg, candidates, candidates_mask)
return gen
def valid(self, req_vars):
valid_dataset = TaxiDataset(self.config.valid_set, 'valid.hdf5')
train_dataset = TaxiDataset('train')
valid_trips_ids = valid_dataset.get_data(
None, slice(0, valid_dataset.num_examples))[
valid_dataset.sources.index('trip_id')]
prefix_stream = DataStream(valid_dataset,
iteration_scheme=SequentialExampleScheme(
valid_dataset.num_examples))
prefix_stream = transformers.taxi_add_datetime(prefix_stream)
prefix_stream = transformers.taxi_add_first_last_len(
prefix_stream, self.config.n_begin_end_pts)
prefix_stream = Batch(prefix_stream,
iteration_scheme=ConstantScheme(
self.config.batch_size))
candidate_stream = DataStream(train_dataset,
iteration_scheme=ShuffledExampleScheme(
train_dataset.num_examples))
candidate_stream = transformers.TaxiExcludeTrips(
candidate_stream, valid_trips_ids)
candidate_stream = transformers.TaxiExcludeEmptyTrips(candidate_stream)
candidate_stream = transformers.taxi_add_datetime(candidate_stream)
candidate_stream = transformers.taxi_add_first_last_len(
candidate_stream, self.config.n_begin_end_pts)
candidate_stream = Batch(candidate_stream,
iteration_scheme=ConstantScheme(
self.config.valid_candidate_size))
sources = prefix_stream.sources + tuple(
'candidate_%s' % k for k in candidate_stream.sources)
stream = Merge((prefix_stream, candidate_stream), sources)
stream = transformers.Select(stream, tuple(req_vars))
stream = MultiProcessing(stream)
return stream
def get_datastream(dataset, batch_size=160):
dataset = DataStream(
dataset,
iteration_scheme=ShuffledScheme(dataset.num_examples, batch_size),
)
dataset = Padding(dataset)
# if flatten:
# dataset = Flatten(dataset, which_sources=('features,'))
def _transpose(data):
return tuple(np.rollaxis(array, 1, 0) for array in data)
dataset = Mapping(dataset, _transpose)
return dataset
def test_padding_data_stream():
# 1-D sequences
stream = Batch(
DataStream(IterableDataset([[1], [2, 3], [], [4, 5, 6], [7]])),
ConstantScheme(2))
mask_stream = Padding(stream)
assert mask_stream.sources == ("data", "data_mask")
it = mask_stream.get_epoch_iterator()
data, mask = next(it)
assert (data == numpy.array([[1, 0], [2, 3]])).all()
assert (mask == numpy.array([[1, 0], [1, 1]])).all()
data, mask = next(it)
assert (data == numpy.array([[0, 0, 0], [4, 5, 6]])).all()
assert (mask == numpy.array([[0, 0, 0], [1, 1, 1]])).all()
data, mask = next(it)
assert (data == numpy.array([[7]])).all()
assert (mask == numpy.array([[1]])).all()
# 2D sequences
stream2 = Batch(
DataStream(IterableDataset([numpy.ones((3, 4)),
2 * numpy.ones((2, 4))])),
ConstantScheme(2))
it = Padding(stream2).get_epoch_iterator()
data, mask = next(it)
assert data.shape == (2, 3, 4)
assert (data[0, :, :] == 1).all()
assert (data[1, :2, :] == 2).all()
assert (mask == numpy.array([[1, 1, 1], [1, 1, 0]])).all()
# 2 sources
stream3 = Padding(Batch(
DataStream(IterableDataset(dict(features=[[1], [2, 3]],
targets=[[4, 5, 6], [7]]))),
ConstantScheme(2)))
assert len(next(stream3.get_epoch_iterator())) == 4
def get_datastream(path, norm_path, which_set='train_si84', batch_size=1):
wsj_dataset = H5PYDataset(path, which_sets=(which_set, ))
data_mean_std = numpy.load(norm_path)
iterator_scheme = ShuffledScheme(batch_size=batch_size,
examples=wsj_dataset.num_examples)
base_stream = DataStream(dataset=wsj_dataset,
iteration_scheme=iterator_scheme)
base_stream = Normalize(data_stream=base_stream,
means=data_mean_std['mean'],
stds=data_mean_std['std'])
fs = FilterSources(data_stream=base_stream,
sources=['features', 'targets'])
padded_stream = Padding(data_stream=fs)
return padded_stream
def setup_sorter_datastream(path, config):
ds = SorterDataset(path)
it = ShuffledExampleScheme(examples=config.example_count)
stream = DataStream(ds, iteration_scheme=it)
stream = Batch(stream,
iteration_scheme=ConstantScheme(config.batch_size *
config.sort_batch_count))
comparison = _balanced_batch_helper(stream.sources.index('unsorted'))
stream = Mapping(stream, SortMapping(comparison))
stream = Unpack(stream)
stream = Batch(stream, iteration_scheme=ConstantScheme(config.batch_size))
stream = Padding(stream,
mask_sources=['answer', 'unsorted'],
mask_dtype='int32')
return ds, stream
def setUp(self):
rng = numpy.random.RandomState(123)
self.stream = DataStream(IndexableDataset(OrderedDict([
('features', rng.rand(4, 2, 2)),
('targets', numpy.array([0, 1, 0, 1]))
]),
axis_labels={
'features':
('batch', 'width',
'height'),
'targets': ('batch', )
}),
iteration_scheme=SequentialScheme(4, 2))
self.duplicate = Duplicate(self.stream, 'features')
def setUp(self):
self.data = OrderedDict([('features', numpy.ones((4, 2, 2))),
('targets',
numpy.array([[0, 1, 2], [1, 0, 1], [1, 1, 1],
[2, 0, 0]]))])
self.neg_data = OrderedDict([('features', numpy.ones((4, 2, 2))),
('targets',
numpy.array([[0, -1, 2], [1, 0, -3],
[1, 1, 1], [2, 0, 0]]))])
self.num_classes = (3, 2, 3)
self.stream_example = StructuredOneHotEncoding(
DataStream(IndexableDataset(self.data),
iteration_scheme=SequentialExampleScheme(4)),
num_classes=self.num_classes,
which_sources=('targets', ))
def _get_sgnmt_dev_stream(val_set=None,
src_vocab=None,
src_vocab_size=30000,
**kwargs):
"""Setup development set stream if necessary.
The arguments to this method are given by the configuration dict.
"""
dev_stream = None
if val_set is not None:
src_vocab = _add_special_ids({str(i) : i
for i in xrange(src_vocab_size)})
dev_dataset = TextFile([val_set], src_vocab, None)
dev_stream = DataStream(dev_dataset)
return dev_stream
def get_test_stream(src_vocab,
trg_vocab,
src_data,
trg_data=None,
src_vocab_size=30000,
unk_id=1,
seq_len=50,
batch_size=80,
sort_k_batches=12,
**kwargs):
"""Prepares the test data stream (=no batches or gold labels)."""
print('streaming...')
# Load dictionaries and ensure special tokens exist
src_vocab = _ensure_special_tokens(src_vocab if isinstance(
src_vocab, dict) else cPickle.load(open(src_vocab)),
bos_idx=0,
eos_idx=2,
unk_idx=unk_id)
trg_vocab = _ensure_special_tokens(trg_vocab if isinstance(
trg_vocab, dict) else cPickle.load(open(trg_vocab)),
bos_idx=0,
eos_idx=2,
unk_idx=unk_id)
# Get text files from both source and target
src_dataset = TextFile([src_data], src_vocab, preprocess=get_unicode)
print(src_data)
#exit()
trg_dataset = TextFile([trg_data], trg_vocab, preprocess=get_unicode)
#stream=DataStream(src_dataset)
stream = Merge([DataStream(src_dataset),
DataStream(trg_dataset)], ('source', 'target'))
return stream
def get_dataset_iterator(dataset,
split,
include_features=True,
include_targets=False,
unit_scale=True,
label_transforms=False,
return_length=False):
"""Get iterator for dataset, split, targets (labels) and scaling (from 255 to 1.0)"""
sources = []
sources = sources + ['features'] if include_features else sources
sources = sources + ['targets'] if include_targets else sources
if split == "all":
splits = ('train', 'valid', 'test')
elif split == "nontrain":
splits = ('valid', 'test')
else:
splits = (split, )
dataset_fname = find_in_data_path("{}.hdf5".format(dataset))
h5_dataset = H5PYDataset(dataset_fname, which_sets=splits, sources=sources)
if unit_scale:
h5_dataset.default_transformers = uint8_pixels_to_floatX(
('features', ))
datastream = DataStream.default_stream(
dataset=h5_dataset,
iteration_scheme=SequentialExampleScheme(h5_dataset.num_examples))
if label_transforms:
# TODO: maybe refactor this common bit with get_custom_streams below
datastream = AddLabelUncertainty(datastream,
chance=0,
which_sources=('targets', ))
datastream = RandomLabelStrip(datastream,
chance=0,
which_sources=('targets', ))
# HACK: allow variable stretch
datastream = StretchLabels(datastream,
length=128,
which_sources=('targets', ))
it = datastream.get_epoch_iterator()
if return_length:
return it, h5_dataset.num_examples
else:
return it
def get_dev_stream(val_set=None,
src_vocab=None,
src_vocab_size=30000,
unk_id=1,
**kwargs):
"""Setup development set stream if necessary."""
dev_stream = None
if val_set is not None and src_vocab is not None:
src_vocab = _ensure_special_tokens(src_vocab if isinstance(
src_vocab, dict) else cPickle.load(open(src_vocab)),
bos_idx=0,
eos_idx=src_vocab_size - 1,
unk_idx=unk_id)
dev_dataset = TextFile([val_set], src_vocab, None)
dev_stream = DataStream(dev_dataset)
return dev_stream
def create_ivector_test_datastream(path, which_set, batch_size=1, delay=0):
wsj_dataset = H5PYDataset(path, which_sets=(which_set, ))
iterator_scheme = SequentialScheme(batch_size=batch_size,
examples=wsj_dataset.num_examples)
base_stream = DataStream(dataset=wsj_dataset,
iteration_scheme=iterator_scheme)
fs = FilterSources(data_stream=base_stream,
sources=['features', 'ivectors', 'targets'])
if delay:
fs = DelayTransformer(fs, delay)
fs = FilterSources(data_stream=fs, sources=['features', 'ivectors'])
return Padding(fs)
def candidate_stream(self, n_candidates):
candidate_stream = DataStream(self.train_dataset,
iteration_scheme=ShuffledExampleScheme(
self.train_dataset.num_examples))
if not data.tvt:
candidate_stream = transformers.TaxiExcludeTrips(
candidate_stream, self.valid_trips_ids)
candidate_stream = transformers.TaxiExcludeEmptyTrips(candidate_stream)
candidate_stream = transformers.taxi_add_datetime(candidate_stream)
candidate_stream = transformers.taxi_add_first_last_len(
candidate_stream, self.config.n_begin_end_pts)
if not data.tvt:
candidate_stream = transformers.add_destination(candidate_stream)
return Batch(candidate_stream,
iteration_scheme=ConstantScheme(n_candidates))
def get_stream(input_file, vocab_file, **kwards):
unk_token = kwards.pop('unk_token')
eos_token = kwards.pop('eos_token')
dataset = TextFile(files=[input_file],
encoding='UTF-8',
preprocess=to_lower_case,
dictionary=pkl.load(open(vocab_file, 'rb')),
level='word',
unk_token=unk_token,
bos_token=None,
eos_token=eos_token)
stream = DataStream(dataset)
return stream
def test_sources_selection():
features = [5, 6, 7, 1]
targets = [1, 0, 1, 1]
stream = DataStream(IterableDataset(OrderedDict(
[('features', features), ('targets', targets)])))
assert list(stream.get_epoch_iterator()) == list(zip(features, targets))
stream = DataStream(IterableDataset(
{'features': features, 'targets': targets},
sources=('targets',)))
assert list(stream.get_epoch_iterator()) == list(zip(targets))
def define_stream(which_sets=('train',),
initial_scale=1,
scale=0.5,
batch_size=64,
seq_length=64,
frame_size=128,
tbptt_flag = True,
num_examples=None):
def _segment_axis(data):
# Defined inside so that frame_size is available
x = tuple([numpy.array([segment_axis(x, frame_size, 0) for x in var])
for var in data])
return x
scale = float(scale)
dataset = Blizzard(which_sets=which_sets)
if num_examples is None:
num_examples = batch_size*(dataset.num_examples/batch_size)
data_stream = DataStream.default_stream(
dataset,
iteration_scheme=SequentialScheme(num_examples, batch_size))
data_stream = ScaleAndShift(data_stream,
scale=1/data_std,
shift=-data_mean/float(data_std))
# Original sampling rate
data_stream = Resample(data_stream, scale=initial_scale)
data_stream = Mapping(data_stream, _copy, add_sources=('upsampled',))
data_stream = Resample(data_stream, scale=scale, which_sources=('upsampled',))
data_stream = Resample(data_stream, scale=1/scale, which_sources=('upsampled',))
# data_stream = Mapping(data_stream, _downsample_and_upsample,
# add_sources=('upsampled',))
data_stream = Mapping(data_stream, _equalize_size)
data_stream = Mapping(data_stream, _get_residual,
add_sources=('residual',))
data_stream = FilterSources(data_stream,
sources=('upsampled', 'residual',))
data_stream = Mapping(data_stream, _segment_axis)
data_stream = Mapping(data_stream, _transpose)
return data_stream
def test_data_stream_mapping_sort_multisource_ndarrays():
data = OrderedDict()
data['x'] = [numpy.array([1, 2, 3]),
numpy.array([2, 3, 1]),
numpy.array([3, 2, 1])]
data['y'] = [numpy.array([6, 5, 4]),
numpy.array([6, 5, 4]),
numpy.array([6, 5, 4])]
data_sorted = [(numpy.array([1, 2, 3]), numpy.array([6, 5, 4])),
(numpy.array([1, 2, 3]), numpy.array([4, 6, 5])),
(numpy.array([1, 2, 3]), numpy.array([4, 5, 6]))]
stream = DataStream(IterableDataset(data))
wrapper = Mapping(stream, mapping=SortMapping(operator.itemgetter(0)))
for output, ground_truth in zip(wrapper.get_epoch_iterator(), data_sorted):
assert len(output) == len(ground_truth)
assert (output[0] == ground_truth[0]).all()
assert (output[1] == ground_truth[1]).all()
def create_data(data):
stream = DataStream(data, iteration_scheme=ShuffledScheme(data.num_examples, batch_size))
# Data Augmentation
stream = MinimumImageDimensions(stream, image_size, which_sources=('image_features',))
stream = MaximumImageDimensions(stream, image_size, which_sources=('image_features',))
stream = RandomHorizontalSwap(stream, which_sources=('image_features',))
stream = Random2DRotation(stream, which_sources=('image_features',))
#stream = ScikitResize(stream, image_size, which_sources=('image_features',))
# Data Preprocessing
# Data Transformation
stream = ScaleAndShift(stream, 1./255, 0, which_sources=('image_features',))
stream = Cast(stream, dtype='float32', which_sources=('image_features',))
return stream
def get_stream(self, part, batch_size=None, max_length=None, seed=None):
dataset = self.get_dataset(part, max_length)
if self._layout == 'lambada' and part == 'train':
stream = DataStream(dataset,
iteration_scheme=RandomSpanScheme(
dataset.num_examples, max_length, seed))
stream = Mapping(stream, listify)
else:
stream = dataset.get_example_stream()
stream = SourcewiseMapping(stream,
functools.partial(add_bos, Vocabulary.BOS))
stream = SourcewiseMapping(stream, vectorize)
if not batch_size:
return stream
stream = Batch(stream, iteration_scheme=ConstantScheme(batch_size))
stream = Padding(stream)
return stream
def get_tst_stream(val_set=None,
src_vocab=None,
src_vocab_size=30000,
unk_id=1,
**kwargs):
tst_stream = None
if val_set is not None and src_vocab is not None:
# Load dictionaries and ensure special tokens exist
src_vocab = ensure_special_tokens(src_vocab if isinstance(
src_vocab, dict) else cPickle.load(open(src_vocab)),
bos_idx=0,
eos_idx=src_vocab_size - 1,
unk_idx=unk_id)
tst_dataset = TextFile([val_set], src_vocab, None)
tst_stream = DataStream(tst_dataset)
return tst_stream
def get_stream(self,
which_set,
shuffle=True,
monitor=False,
num_examples=None,
center=True):
scheme_klass = ShuffledScheme if shuffle else SequentialScheme
if num_examples is None:
num_examples = self.get_stream_num_examples(which_set,
monitor=monitor)
scheme = scheme_klass(num_examples, self.batch_size)
stream = DataStream.default_stream(dataset=self.datasets[which_set],
iteration_scheme=scheme)
stream = self.apply_default_transformers(stream)
stream = Canonicalize(stream, mapping=self.preprocess)
if center:
stream = transformers.Mapping(stream, mapping=self.center)
return stream
def stream_handwriting(which_sets, batch_size, seq_size, tbptt=True):
dataset = Handwriting(which_sets)
data_stream = DataStream.default_stream(
dataset,
iteration_scheme=ShuffledScheme(
batch_size * (dataset.num_examples / batch_size), batch_size))
data_stream = FilterSources(data_stream, sources=('features', ))
data_stream = Padding(data_stream)
data_stream = Mapping(data_stream, _transpose)
if tbptt:
data_stream = SegmentSequence(data_stream,
add_flag=True,
seq_size=seq_size)
data_stream = ForceFloatX(data_stream)
return data_stream
def start_server(port, which_set):
fuel.server.logger.setLevel('WARN')
dataset = IMDBText(which_set, sorted=True)
n_train = dataset.num_examples
#scheme = ShuffledScheme(examples=n_train, batch_size=batch_size)
scheme = BatchwiseShuffledScheme(examples=n_train, batch_size=batch_size)
stream = DataStream(
dataset=dataset,
iteration_scheme=scheme)
print "loading glove"
glove = GloveTransformer(glove_version, data_stream=stream)
padded = Padding(
data_stream=glove,
mask_sources=('features',)
)
fuel.server.start_server(padded, port=port, hwm=20)
def get_comb_stream(fea2obj, which_set, batch_size=None, shuffle=True):
streams = []
for fea in fea2obj:
obj = fea2obj[fea]
dataset = H5PYDataset(obj.fuelfile, which_sets=(which_set,),load_in_memory=True)
if batch_size == None: batch_size = dataset.num_examples
if shuffle:
iterschema = ShuffledScheme(examples=dataset.num_examples, batch_size=batch_size)
else:
iterschema = SequentialScheme(examples=dataset.num_examples, batch_size=batch_size)
stream = DataStream(dataset=dataset, iteration_scheme=iterschema)
if fea in seq_features:
stream = CutInput(stream, obj.max_len)
if obj.rec == True:
logger.info('transforming data for recursive input')
stream = LettersTransposer(stream, which_sources=fea)# Required because Recurrent bricks receive as input [sequence, batch,# features]
streams.append(stream)
stream = Merge(streams, tuple(fea2obj.keys()))
return stream, dataset.num_examples
def get_dev_streams(config):
"""Setup development set stream if necessary."""
dev_streams = {}
for cg in config['cgs']:
if 'val_sets' in config and cg in config['val_sets']:
logger.info('Building development stream for cg:[{}]'.format(cg))
eid = p_(cg)[0]
dev_file = config['val_sets'][cg]
# Get dictionary and fix EOS
dictionary = cPickle.load(open(config['src_vocabs'][eid]))
dictionary['<S>'] = 0
dictionary['<UNK>'] = config['unk_id']
dictionary['</S>'] = config['src_eos_idxs'][eid]
# Get as a text file and convert it into a stream
dev_dataset = TextFile([dev_file], dictionary, None)
dev_streams[cg] = DataStream(dev_dataset)
return dev_streams
def create_data(data):
stream = DataStream.default_stream(data,
iteration_scheme=ShuffledScheme(
data.num_examples, batch_size))
stream_downscale = MinimumImageDimensions(
stream, image_size, which_sources=('image_features', ))
#stream_rotate = Random2DRotation(stream_downscale, which_sources=('image_features',))
stream_max = ScikitResize(stream_downscale,
image_size,
which_sources=('image_features', ))
stream_scale = ScaleAndShift(stream_max,
1. / 255,
0,
which_sources=('image_features', ))
stream_cast = Cast(stream_scale,
dtype='float32',
which_sources=('image_features', ))
#stream_flat = Flatten(stream_scale, which_sources=('image_features',))
return stream_cast