def __init__(self,
data=None,
data_specs=None,
rng=_default_seed,
preprocessor=None,
fit_preprocessor=False):
# data_specs should be flat, and there should be no
# duplicates in source, as we keep only one version
assert is_flat_specs(data_specs)
if isinstance(data_specs[1], tuple):
assert sorted(set(data_specs[1])) == sorted(data_specs[1])
space, source = data_specs
space.np_validate(data)
assert len(set(elem.shape[0] for elem in list(data))) <= 1
self.data = data
self.data_specs = data_specs
self.num_examples = list(data)[0].shape[0]
self.compress = False
self.design_loc = None
self.rng = make_np_rng(rng, which_method='random_integers')
# Defaults for iterators
self._iter_mode = resolve_iterator_class('sequential')
if preprocessor:
preprocessor.apply(self, can_fit=fit_preprocessor)
self.preprocessor = preprocessor
def __init__(self, data=None, data_specs=None, rng=_default_seed,
preprocessor=None, fit_preprocessor=False):
# data_specs should be flat, and there should be no
# duplicates in source, as we keep only one version
assert is_flat_specs(data_specs)
if isinstance(data_specs[1], tuple):
assert sorted(set(data_specs[1])) == sorted(data_specs[1])
space, source = data_specs
space.np_validate(data)
# TODO: assume that data[0] is num example => error if channel in c01b
# assert len(set(elem.shape[0] for elem in list(data))) <= 1
self.data = data
self.data_specs = data_specs
# TODO: assume that data[0] is num example => error if channel in c01b
self.num_examples = list(data)[-1].shape[0] # TODO: list(data)[0].shape[0]
self.compress = False
self.design_loc = None
self.rng = make_np_rng(rng, which_method='random_integers')
# Defaults for iterators
self._iter_mode = resolve_iterator_class('sequential')
if preprocessor:
preprocessor.apply(self, can_fit=fit_preprocessor)
self.preprocessor = preprocessor
def iterator(self,
mode=None,
batch_size=None,
num_batches=None,
rng=None,
data_specs=None,
return_tuple=False):
"""
.. todo::
WRITEME
"""
# Build the right data_specs to query self.raw
if data_specs is not None:
assert is_flat_specs(data_specs)
space, source = data_specs
if not isinstance(source, tuple):
source = (source, )
if isinstance(space, CompositeSpace):
space = tuple(space.components)
else:
space = (space, )
# Put 'features' first, as this is what TransformerIterator
# is expecting
if 'features' not in source:
# 'features is not needed, get things directly from
# the original data
raw_data_specs = data_specs
else:
feature_idx = source.index('features')
if self.space_preserving:
# Ask self.raw for the data in the expected space,
# and self.transformer will operate in that space
feature_input_space = space[feature_idx]
else:
# We need to ask the transformer what its input space is
feature_input_space = self.transformer.get_input_space()
raw_space = CompositeSpace((feature_input_space, ) +
space[:feature_idx] +
space[feature_idx + 1:])
raw_source = (('features', ) + source[:feature_idx] +
source[feature_idx + 1:])
raw_data_specs = (raw_space, raw_source)
else:
raw_data_specs = None
raw_iterator = self.raw.iterator(mode=mode,
batch_size=batch_size,
num_batches=num_batches,
rng=rng,
data_specs=raw_data_specs,
return_tuple=return_tuple)
final_iterator = TransformerIterator(raw_iterator,
self,
data_specs=data_specs)
return final_iterator
def iterator(self, mode=None, batch_size=None, num_batches=None,
topo=None, targets=None, rng=None, data_specs=None,
return_tuple=False):
"""
.. todo::
WRITEME
"""
# Build the right data_specs to query self.raw
if data_specs is not None:
assert is_flat_specs(data_specs)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if isinstance(space, CompositeSpace):
space = tuple(space.components)
else:
space = (space,)
# Put 'features' first, as this is what TransformerIterator
# is expecting
if 'features' not in source:
# 'features is not needed, get things directly from
# the original data
raw_data_specs = data_specs
else:
feature_idx = source.index('features')
if self.space_preserving:
# Ask self.raw for the data in the expected space,
# and self.transformer will operate in that space
feature_input_space = space[feature_idx]
else:
# We need to ask the transformer what its input space is
feature_input_space = self.transformer.get_input_space()
raw_space = CompositeSpace((feature_input_space,)
+ space[:feature_idx]
+ space[feature_idx + 1:])
raw_source = (('features',)
+ source[:feature_idx]
+ source[feature_idx + 1:])
raw_data_specs = (raw_space, raw_source)
else:
raw_data_specs = None
raw_iterator = self.raw.iterator(
mode=mode, batch_size=batch_size,
num_batches=num_batches, topo=topo, targets=targets, rng=rng,
data_specs=raw_data_specs, return_tuple=return_tuple)
final_iterator = TransformerIterator(raw_iterator, self,
data_specs=data_specs)
return final_iterator
def __init__(self,
data=None,
data_specs=None,
rng=_default_seed,
preprocessor=None,
fit_preprocessor=False):
"""
Parameters
----------
data: ndarray, or tuple of ndarrays, containing the data.
It is formatted as specified in `data_specs`.
For instance, if `data_specs` is (VectorSpace(nfeat), 'features'),
then `data` has to be a 2-d ndarray, of shape (nb examples,
nfeat), that defines an unlabeled dataset. If `data_specs`
is (CompositeSpace(Conv2DSpace(...), VectorSpace(1)),
('features', 'target')), then `data` has to be an (X, y) pair,
with X being an ndarray containing images stored in the topological
view specified by the `Conv2DSpace`, and y being a 2-D ndarray
of width 1, containing the labels or targets for each example.
data_specs: A (space, source) pair, where space is an instance of
`Space` (possibly a `CompositeSpace`), and `source` is a
string (or tuple of strings, if `space` is a `CompositeSpace`),
defining the format and labels associated to `data`.
rng : object, optional
A random number generator used for picking random
indices into the design matrix when choosing minibatches.
preprocessor: WRITEME
fit_preprocessor: WRITEME
"""
# data_specs should be flat, and there should be no
# duplicates in source, as we keep only one version
assert is_flat_specs(data_specs)
if isinstance(data_specs[1], tuple):
assert sorted(set(data_specs[1])) == sorted(data_specs[1])
self.data = data
self.data_specs = data_specs
self.compress = False
self.design_loc = None
if hasattr(rng, 'random_integers'):
self.rng = rng
else:
self.rng = np.random.RandomState(rng)
# Defaults for iterators
self._iter_mode = resolve_iterator_class('sequential')
if preprocessor:
preprocessor.apply(self, can_fit=fit_preprocessor)
self.preprocessor = preprocessor
def __init__(self, data=None, data_specs=None, rng=_default_seed,
preprocessor=None, fit_preprocessor=False):
"""
Parameters
----------
data: ndarray, or tuple of ndarrays, containing the data.
It is formatted as specified in `data_specs`.
For instance, if `data_specs` is (VectorSpace(nfeat), 'features'),
then `data` has to be a 2-d ndarray, of shape (nb examples,
nfeat), that defines an unlabeled dataset. If `data_specs`
is (CompositeSpace(Conv2DSpace(...), VectorSpace(1)),
('features', 'target')), then `data` has to be an (X, y) pair,
with X being an ndarray containing images stored in the topological
view specified by the `Conv2DSpace`, and y being a 2-D ndarray
of width 1, containing the labels or targets for each example.
data_specs: A (space, source) pair, where space is an instance of
`Space` (possibly a `CompositeSpace`), and `source` is a
string (or tuple of strings, if `space` is a `CompositeSpace`),
defining the format and labels associated to `data`.
rng : object, optional
A random number generator used for picking random
indices into the design matrix when choosing minibatches.
preprocessor: WRITEME
fit_preprocessor: WRITEME
"""
# data_specs should be flat, and there should be no
# duplicates in source, as we keep only one version
assert is_flat_specs(data_specs)
if isinstance(data_specs[1], tuple):
assert sorted(set(data_specs[1])) == sorted(data_specs[1])
self.data = data
self.data_specs = data_specs
self.compress = False
self.design_loc = None
if hasattr(rng, 'random_integers'):
self.rng = rng
else:
self.rng = np.random.RandomState(rng)
# Defaults for iterators
self._iter_mode = resolve_iterator_class('sequential')
if preprocessor:
preprocessor.apply(self, can_fit=fit_preprocessor)
self.preprocessor = preprocessor
def iterator(self, mode=None, batch_size=None, num_batches=None,
rng=None, data_specs=None,
return_tuple=False):
"""
.. todo::
WRITEME
"""
# Build the right data_specs to query self.raw
feature_idx = None
if data_specs is not None:
assert is_flat_specs(data_specs)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if isinstance(space, CompositeSpace):
space = tuple(space.components)
else:
space = (space,)
if self.transform_source not in source:
# 'features is not needed, get things directly from
# the original data
raw_data_specs = data_specs
else:
feature_idx = source.index(self.transform_source)
feature_input_space = self.transformer.get_input_space()
raw_space = CompositeSpace(space[:feature_idx]
+ (feature_input_space,)
+ space[feature_idx + 1:])
raw_source = source
raw_data_specs = (raw_space, raw_source)
else:
raw_data_specs = None
raw_iterator = self.raw.iterator(
mode=mode, batch_size=batch_size,
num_batches=num_batches, rng=rng,
data_specs=raw_data_specs, return_tuple=return_tuple)
final_iterator = TransformerIterator(raw_iterator, self,
feature_idx=feature_idx,
data_specs=data_specs)
return final_iterator
def __init__(self, data=None, data_specs=None, rng=_default_seed,
preprocessor=None, fit_preprocessor=False):
# data_specs should be flat, and there should be no
# duplicates in source, as we keep only one version
assert is_flat_specs(data_specs)
if isinstance(data_specs[1], tuple):
assert sorted(set(data_specs[1])) == sorted(data_specs[1])
self.data = data
self.data_specs = data_specs
self.compress = False
self.design_loc = None
self.rng = make_np_rng(rng, which_method='random_integers')
# Defaults for iterators
self._iter_mode = resolve_iterator_class('sequential')
if preprocessor:
preprocessor.apply(self, can_fit=fit_preprocessor)
self.preprocessor = preprocessor
def iterator(self, mode=None, batch_size=None, num_batches=None,
topo=None, targets=None, rng=None, data_specs=None,
return_tuple=False):
# Build the right data_specs to query self.raw
if data_specs is not None:
assert is_flat_specs(data_specs)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if isinstance(space, CompositeSpace):
space = tuple(space.components)
else:
space = (space,)
# Put 'features' first, as this is what TransformerIterator
# is expecting
if 'features' not in source:
# 'features is not needed, get things directly from
# the original data
raw_data_specs = data_specs
else:
feature_idx = source.index('features')
raw_space = CompositeSpace(
(self.transformer.get_input_space(),)
+ space[:feature_idx]
+ space[feature_idx + 1:])
raw_source = (('features',)
+ source[:feature_idx]
+ source[feature_idx + 1:])
raw_data_specs = (raw_space, raw_source)
else:
raw_data_specs = None
raw_iterator = self.raw.iterator(mode=mode, batch_size=batch_size,
num_batches=num_batches, topo=topo, targets=targets, rng=rng,
data_specs=raw_data_specs, return_tuple=return_tuple)
final_iterator = TransformerIterator(raw_iterator, self,
data_specs=data_specs)
return final_iterator
def __init__(self, dataset, subset_iterator, topo=None, targets=None,
data_specs=None, return_tuple=False):
if topo is not None or targets is not None:
if data_specs is not None:
raise ValueError("In FiniteDatasetIterator, both "
"the `data_specs` argument and deprecated arguments "
"`topo` or `targets` were provided.",
(data_specs, topo, targets))
warnings.warn("Usage of `topo` and `target` arguments are being "
"deprecated, and will be removed around November 7th, "
"2013. `data_specs` should be used instead.",
stacklevel=2)
topo = False
targets = False
self._deprecated_interface = True
else:
self._deprecated_interface = False
self._topo = topo
self._targets = targets
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# TODO: More thought about how to handle things where this
# fails (gigantic HDF5 files, etc.)
if self._deprecated_interface:
self._raw_data = self._dataset.get_design_matrix()
if self._targets:
self._raw_targets = self._dataset.get_targets()
if self._raw_targets is None:
raise ValueError("Can't iterate with targets=True on a "
"dataset object with no targets")
self._targets_need_cast = not np.dtype(config.floatX) == self._raw_targets.dtype
self._needs_cast = not np.dtype(config.floatX) == self._raw_data.dtype
else:
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data,)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._raw_data = tuple(all_data[dataset_source.index(s)]
for s in source)
self._source = source
self._convert = []
for i, (so, sp) in enumerate(safe_zip(source, sub_spaces)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
# Compose the functions
fn = None
needs_cast = not np.dtype(config.floatX) == \
self._raw_data[i].dtype
if needs_cast:
fn = lambda batch: numpy.cast[config.floatX](batch)
needs_format = not sp == dspace
if needs_format:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
fn = (lambda batch, dspace=dspace, sp=sp:
dspace.np_format_as(batch, sp))
else:
fn = (lambda batch, dspace=dspace, sp=sp, fn_=fn:
dspace.np_format_as(fn_(batch), sp))
self._convert.append(fn)
def __init__(self, dataset, subset_iterator, data_specs=None,
return_tuple=False, convert=None):
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, (tuple, list)):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
# If `dataset` is incompatible with the new interface, fall back to the
# old interface
if not hasattr(self._dataset, 'get'):
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data,)
raw_data = []
for s in source:
try:
raw_data.append(all_data[dataset_source.index(s)])
except ValueError as e:
msg = str(e) + '\nThe dataset does not provide '\
'a source with name: ' + s + '.'
reraise_as(ValueError(msg))
self._raw_data = tuple(raw_data)
self._source = source
self._space = sub_spaces
if convert is None:
self._convert = [None for s in source]
else:
assert len(convert) == len(source)
self._convert = convert
for i, (so, sp) in enumerate(safe_izip(source, sub_spaces)):
try:
idx = dataset_source.index(so)
except ValueError as e:
msg = str(e) + '\nThe dataset does not provide '\
'a source with name: ' + so + '.'
reraise_as(ValueError(msg))
dspace = dataset_sub_spaces[idx]
fn = self._convert[i]
# If there is a fn, it is supposed to take care of the formatting,
# and it should be an error if it does not. If there was no fn,
# then the iterator will try to format using the generic
# space-formatting functions.
if fn is None:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
fn = (lambda batch, dspace=dspace, sp=sp:
dspace.np_format_as(batch, sp))
self._convert[i] = fn
def __init__(self,
dataset,
subset_iterator,
data_specs=None,
return_tuple=False,
convert=None):
"""
.. todo::
WRITEME
"""
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source, )
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space, )
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
space, source = data_specs
if not isinstance(source, tuple):
source = (source, )
if not isinstance(space, CompositeSpace):
sub_spaces = (space, )
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._source = source
if convert is None:
self._convert = [None for s in source]
else:
assert len(convert) == len(source)
self._convert = convert
for i, (so, sp) in enumerate(safe_zip(source, sub_spaces)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
init_fn = self._convert[i]
fn = init_fn
# If there is an init_fn, it is supposed to take
# care of the formatting, and it should be an error
# if it does not. If there was no init_fn, then
# the iterator will try to format using the generic
# space-formatting functions.
if init_fn is None:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
fn = (lambda batch, dspace=dspace, sp=sp: dspace.
np_format_as(batch, sp))
else:
fn = (lambda batch, dspace=dspace, sp=sp, fn_=fn: dspace.
np_format_as(fn_(batch), sp))
self._convert[i] = fn
def __init__(self,
dataset,
subset_iterator,
data_specs=None,
return_tuple=False,
convert=None):
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, (tuple, list)):
dataset_source = (dataset_source, )
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space, )
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
space, source = data_specs
if not isinstance(source, tuple):
source = (source, )
if not isinstance(space, CompositeSpace):
sub_spaces = (space, )
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
# If `dataset` is incompatible with the new interface, fall back to the
# old interface
if not hasattr(self._dataset, 'get'):
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data, )
raw_data = []
for s in source:
try:
raw_data.append(all_data[dataset_source.index(s)])
except ValueError as e:
msg = str(e) + '\nThe dataset does not provide '\
'a source with name: ' + s + '.'
reraise_as(ValueError(msg))
self._raw_data = tuple(raw_data)
self._source = source
self._space = sub_spaces
if convert is None:
self._convert = [None for s in source]
else:
assert len(convert) == len(source)
self._convert = convert
for i, (so, sp) in enumerate(safe_izip(source, sub_spaces)):
try:
idx = dataset_source.index(so)
except ValueError as e:
msg = str(e) + '\nThe dataset does not provide '\
'a source with name: ' + so + '.'
reraise_as(ValueError(msg))
dspace = dataset_sub_spaces[idx]
fn = self._convert[i]
# If there is a fn, it is supposed to take care of the formatting,
# and it should be an error if it does not. If there was no fn,
# then the iterator will try to format using the generic
# space-formatting functions.
if fn is None:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
fn = (lambda batch, dspace=dspace, sp=sp: dspace.np_format_as(
batch, sp))
self._convert[i] = fn
def __init__(self, dataset, subset_iterator, data_specs=None,
return_tuple=False, convert=None):
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._source = source
if convert is None:
self._convert = [None for s in source]
else:
assert len(convert) == len(source)
self._convert = convert
for i, (so, sp) in enumerate(safe_izip(source, sub_spaces)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
init_fn = self._convert[i]
fn = init_fn
# If there is an init_fn, it is supposed to take
# care of the formatting, and it should be an error
# if it does not. If there was no init_fn, then
# the iterator will try to format using the generic
# space-formatting functions.
if init_fn is None:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
fn = (lambda batch, dspace=dspace, sp=sp:
dspace.np_format_as(batch, sp))
else:
fn = (lambda batch, dspace=dspace, sp=sp, fn_=fn:
dspace.np_format_as(fn_(batch), sp))
self._convert[i] = fn
def __init__(self, dataset, subset_iterator, topo=None, targets=None,
data_specs=None, return_tuple=False):
if topo is not None or targets is not None:
if data_specs is not None:
raise ValueError("In FiniteDatasetIterator, both "
"the `data_specs` argument and deprecated arguments "
"`topo` or `targets` were provided.",
(data_specs, topo, targets))
warnings.warn("Usage of `topo` and `target` arguments are being "
"deprecated, and will be removed around November 7th, "
"2013. `data_specs` should be used instead.",
stacklevel=2)
topo = False
targets = False
self._deprecated_interface = True
else:
self._deprecated_interface = False
self._topo = topo
self._targets = targets
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# TODO: More thought about how to handle things where this
# fails (gigantic HDF5 files, etc.)
if self._deprecated_interface:
self._raw_data = self._dataset.get_design_matrix()
if self._targets:
self._raw_targets = self._dataset.get_targets()
if self._raw_targets is None:
raise ValueError("Can't iterate with targets=True on a "
"dataset object with no targets")
self._targets_need_cast = not np.dtype(config.floatX) == self._raw_targets.dtype
self._needs_cast = not np.dtype(config.floatX) == self._raw_data.dtype
else:
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data,)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._raw_data = tuple(all_data[dataset_source.index(s)]
for s in source)
self._source = source
self._convert = []
for i, (so, sp) in enumerate(safe_zip(source, sub_spaces)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
# Compose the functions
fn = None
needs_cast = not np.dtype(config.floatX) == \
self._raw_data[i].dtype
if needs_cast:
fn = lambda batch: numpy.cast[config.floatX](batch)
needs_format = not sp == dspace
if needs_format:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
fn = (lambda batch, dspace=dspace, sp=sp:
dspace.np_format_as(batch, sp))
else:
fn = (lambda batch, dspace=dspace, sp=sp:
dspace.np_format_as(fn(batch), sp))
self._convert.append(fn)
def __init__(self, dataset, subset_iterator, data_specs=None,
return_tuple=False, convert=None):
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data,)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._raw_data = tuple(all_data[dataset_source.index(s)]
for s in source)
self._source = source
if convert is None:
self._convert = [None for s in source]
else:
assert len(convert) == len(source)
self._convert = convert
for i, (so, sp, dt) in enumerate(safe_izip(source,
sub_spaces,
self._raw_data)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
init_fn = self._convert[i]
fn = init_fn
# If there is an init_fn, it is supposed to take
# care of the formatting, and it should be an error
# if it does not. If there was no init_fn, then
# the iterator will try to format using the generic
# space-formatting functions.
if init_fn is None:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
def fn(batch, dspace=dspace, sp=sp):
try:
return dspace.np_format_as(batch, sp)
except ValueError as e:
msg = str(e) + '\nMake sure that the model and '\
'dataset have been initialized with '\
'correct values.'
reraise_as(ValueError(msg))
else:
fn = (lambda batch, dspace=dspace, sp=sp, fn_=fn:
dspace.np_format_as(fn_(batch), sp))
self._convert[i] = fn
def __init__(self, dataset, subset_iterator, data_specs=None, return_tuple=False, convert_fns=None):
"""
convert_fns: function or tuple of function, organized as
in data_specs, to be applied on the raw batches of
data. "None" can be used as placeholder for the identity.
"""
self._deprecated_interface = False
if data_specs is None:
raise TypeError("data_specs not provided")
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data,)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._raw_data = tuple(all_data[dataset_source.index(s)] for s in source)
self._source = source
if convert_fns is None:
self._convert = [None for s in source]
else:
if not isinstance(convert_fns, (list, tuple)):
convert_fns = (convert_fns,)
assert len(convert_fns) == len(source)
self._convert = list(convert_fns)
for i, (so, sp) in enumerate(safe_zip(source, sub_spaces)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
# Compose the functions
fn = self._convert[i]
needs_cast = not (self._raw_data[i][0].dtype == config.floatX)
if needs_cast:
if fn is None:
fn = lambda batch: np.cast[config.floatX](batch)
else:
fn = lambda batch, prev_fn=fn: np.cast[config.floatX](prev_fn(batch))
needs_format = not sp == dspace
if needs_format:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
fn = lambda batch, dspace=dspace, sp=sp: dspace.np_format_as(batch, sp)
else:
fn = lambda batch, dspace=dspace, sp=sp, fn_=fn: dspace.np_format_as(fn_(batch), sp)
self._convert[i] = fn
def __init__(self, dataset, subset_iterator, data_specs=None,
return_tuple=False, convert=None):
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data,)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._raw_data = ()
for s in source:
try:
self._raw_data += (all_data[dataset_source.index(s)],)
except ValueError as e:
msg = str(e) + '\nThe dataset does not provide '\
'a source with name: '+s+'.'
reraise_as(ValueError(msg))
self._source = source
self._space = sub_spaces
if convert is None:
self._convert = [None for s in source]
else:
assert len(convert) == len(source)
self._convert = convert
for i, (so, sp, dt) in enumerate(safe_izip(source,
sub_spaces,
self._raw_data)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
init_fn = self._convert[i]
fn = init_fn
# If there is an init_fn, it is supposed to take
# care of the formatting, and it should be an error
# if it does not. If there was no init_fn, then
# the iterator will try to format using the generic
# space-formatting functions.
if init_fn is None:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
def fn(batch, dspace=dspace, sp=sp):
try:
return dspace.np_format_as(batch, sp)
except ValueError as e:
msg = str(e) + '\nMake sure that the model and '\
'dataset have been initialized with '\
'correct values.'
reraise_as(ValueError(msg))
else:
fn = (lambda batch, dspace=dspace, sp=sp, fn_=fn:
dspace.np_format_as(fn_(batch), sp))
self._convert[i] = fn
def __init__(self,
dataset,
subset_iterator,
data_specs=None,
return_tuple=False,
convert_fns=None):
"""
convert_fns: function or tuple of function, organized as
in data_specs, to be applied on the raw batches of
data. "None" can be used as placeholder for the identity.
"""
self._deprecated_interface = False
if data_specs is None:
raise TypeError("data_specs not provided")
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
# Keep only the needed sources in self._raw_data.
# Remember what source they correspond to in self._source
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source, )
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space, )
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
all_data = self._dataset.get_data()
if not isinstance(all_data, tuple):
all_data = (all_data, )
space, source = data_specs
if not isinstance(source, tuple):
source = (source, )
if not isinstance(space, CompositeSpace):
sub_spaces = (space, )
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._raw_data = tuple(all_data[dataset_source.index(s)]
for s in source)
self._source = source
if convert_fns is None:
self._convert = [None for s in source]
else:
if not isinstance(convert_fns, (list, tuple)):
convert_fns = (convert_fns, )
assert len(convert_fns) == len(source)
self._convert = list(convert_fns)
for i, (so, sp) in enumerate(safe_zip(source, sub_spaces)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
# Compose the functions
fn = self._convert[i]
needs_cast = not (self._raw_data[i][0].dtype == config.floatX)
if needs_cast:
if fn is None:
fn = lambda batch: np.cast[config.floatX](batch)
else:
fn = (lambda batch, prev_fn=fn: np.cast[config.floatX]
(prev_fn(batch)))
needs_format = not sp == dspace
if needs_format:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
fn = (lambda batch, dspace=dspace, sp=sp: dspace.
np_format_as(batch, sp))
else:
fn = (lambda batch, dspace=dspace, sp=sp, fn_=fn: dspace.
np_format_as(fn_(batch), sp))
self._convert[i] = fn
def __init__(self, dataset, subset_iterator,
data_specs=None, return_tuple=False, convert=None):
"""
.. todo::
WRITEME
"""
self._data_specs = data_specs
self._dataset = dataset
self._subset_iterator = subset_iterator
self._return_tuple = return_tuple
assert is_flat_specs(data_specs)
dataset_space, dataset_source = self._dataset.get_data_specs()
assert is_flat_specs((dataset_space, dataset_source))
# the dataset's data spec is either a single (space, source) pair,
# or a pair of (non-nested CompositeSpace, non-nested tuple).
# We could build a mapping and call flatten(..., return_tuple=True)
# but simply putting spaces, sources and data in tuples is simpler.
if not isinstance(dataset_source, tuple):
dataset_source = (dataset_source,)
if not isinstance(dataset_space, CompositeSpace):
dataset_sub_spaces = (dataset_space,)
else:
dataset_sub_spaces = dataset_space.components
assert len(dataset_source) == len(dataset_sub_spaces)
space, source = data_specs
if not isinstance(source, tuple):
source = (source,)
if not isinstance(space, CompositeSpace):
sub_spaces = (space,)
else:
sub_spaces = space.components
assert len(source) == len(sub_spaces)
self._source = source
if convert is None:
self._convert = [None for s in source]
else:
assert len(convert) == len(source)
self._convert = convert
dtypes = self._dataset.dtype_of(self._source)
for i, (so, sp, dt) in enumerate(safe_zip(source, sub_spaces, dtypes)):
idx = dataset_source.index(so)
dspace = dataset_sub_spaces[idx]
init_fn = self._convert[i]
fn = init_fn
# Compose the functions
needs_cast = not (numpy.dtype(config.floatX) == dt)
if needs_cast:
if fn is None:
fn = lambda batch: numpy.cast[config.floatX](batch)
else:
fn = (lambda batch, fn_=fn:
numpy.cast[config.floatX](fn_(batch)))
# If there is an init_fn, it is supposed to take
# care of the formatting, and it should be an error
# if it does not. If there was no init_fn, then
# the iterator will try to format using the generic
# space-formatting functions.
needs_format = not init_fn and not sp == dspace
if needs_format:
# "dspace" and "sp" have to be passed as parameters
# to lambda, in order to capture their current value,
# otherwise they would change in the next iteration
# of the loop.
if fn is None:
fn = (lambda batch, dspace=dspace, sp=sp:
dspace.np_format_as(batch, sp))
else:
fn = (lambda batch, dspace=dspace, sp=sp, fn_=fn:
dspace.np_format_as(fn_(batch), sp))
self._convert[i] = fn