def next(self):
next_index = self._subset_iterator.next()
if isinstance(next_index, numpy.ndarray) and len(next_index) == 1:
next_index = next_index[0]
self._needs_cast = True
if self._needs_cast:
features = numpy.cast[config.floatX](self._raw_data[0][next_index])
else:
features = self._raw_data[0][next_index, :]
#import ipdb; ipdb.set_trace()
if self.topo:
if len(features.shape) != 2:
features = features.reshape((1, features.shape[0]))
features = self._dataset.get_topological_view(features)
if self.targets:
bbx_targets = get_image_bboxes(next_index, self._raw_data[1])
if len(bbx_targets.shape) != 2:
bbx_targets = bbx_targets.reshape((1, bbx_targets.shape[0]))
if self.use_output_map:
if self.bbox_conversion_type == ConversionType.GUID:
if isinstance(self._data_specs[0].components[1],
Conv2DSpace):
conv_outs = True
else:
conv_outs = False
n_channels = self._data_specs[0].components[1].num_channels
targets = convert_bboxes_guided(bbx_targets,
self.img_shape,
self.receptive_field_shape,
self.area_ratio,
self.stride,
conv_outs=conv_outs,
n_channels=n_channels)
else:
targets = convert_bboxes_exhaustive(
bbx_targets, self.img_shape,
self.receptive_field_shape, self.area_ratio,
self.stride)
else:
targets = self.get_bare_outputs(bbx_targets)
if targets.shape[0] != features.shape[0]:
raise ValueError(
"There is a batch size mismatch between features and targets."
)
self._targets_need_cast = True
if self._targets_need_cast:
targets = numpy.cast[config.floatX](targets)
return features, targets
else:
if self._return_tuple:
features = (features, )
return features
def resize(h5file, start, stop, remove_old_node=False):
if h5file is None:
raise ValueError("h5file should not be None.")
data = h5file.root.Data
node_name = "Data_%s_%s" % (start, stop)
if remove_old_node:
try:
gcolumns = h5file.createGroup('/', node_name,
"Data %s" % node_name)
except tables.exceptions.NodeError:
h5file.removeNode('/', node_name, 1)
gcolumns = h5file.createGroup('/', node_name,
"Data %s" % node_name)
elif node_name in h5file.root:
return h5file, getattr(h5file.root, node_name)
else:
gcolumns = h5file.createGroup('/', node_name,
"Data %s" % node_name)
if FaceBBoxDDMPytables.h5file is None:
FaceBBoxDDMPytables.h5file = h5file
start = 0 if start is None else start
stop = gcolumns.X.nrows if stop is None else stop
atom = tables.Float32Atom(
) if config.floatX == 'float32' else tables.Float64Atom()
filters = FaceBBoxDDMPytables.filters
x = h5file.createCArray(gcolumns,
'X',
atom=atom,
shape=((stop - start, data.X.shape[1])),
title="Images",
filters=filters)
y = h5file.createTable(gcolumns,
'bboxes',
BoundingBox,
title="Face bounding boxes",
filters=filters)
x[:] = data.X[start:stop]
bboxes = get_image_bboxes(slice(start, stop), data.bboxes)
y.append(bboxes)
if remove_old_node:
h5file.removeNode('/', "Data", 1)
h5file.renameNode('/', "Data", node_name)
h5file.flush()
return h5file, gcolumns
def next(self):
next_index = self._subset_iterator.next()
if isinstance(next_index, numpy.ndarray) and len(next_index) == 1:
next_index = next_index[0]
self._needs_cast = True
if self._needs_cast:
features = numpy.cast[config.floatX](self._raw_data[0][next_index])
else:
features = self._raw_data[0][next_index,:]
#import ipdb; ipdb.set_trace()
if self.topo:
if len(features.shape) != 2:
features = features.reshape((1, features.shape[0]))
features = self._dataset.get_topological_view(features)
if self.targets:
bbx_targets = get_image_bboxes(next_index, self._raw_data[1])
if len(bbx_targets.shape) != 2:
bbx_targets = bbx_targets.reshape((1, bbx_targets.shape[0]))
if self.use_output_map:
if self.bbox_conversion_type == ConversionType.GUID:
if isinstance(self._data_specs[0].components[1], Conv2DSpace):
conv_outs = True
else:
conv_outs = False
n_channels = self._data_specs[0].components[1].num_channels
targets = convert_bboxes_guided(bbx_targets, self.img_shape,
self.receptive_field_shape, self.area_ratio, self.stride,
conv_outs=conv_outs, n_channels=n_channels)
else:
targets = convert_bboxes_exhaustive(bbx_targets, self.img_shape,
self.receptive_field_shape, self.area_ratio, self.stride)
else:
targets = self.get_bare_outputs(bbx_targets)
if targets.shape[0] != features.shape[0]:
raise ValueError("There is a batch size mismatch between features and targets.")
self._targets_need_cast = True
if self._targets_need_cast:
targets = numpy.cast[config.floatX](targets)
return features, targets
else:
if self._return_tuple:
features = (features,)
return features
def resize(h5file, start, stop, remove_old_node=False):
if h5file is None:
raise ValueError("h5file should not be None.")
data = h5file.root.Data
node_name = "Data_%s_%s" % (start, stop)
if remove_old_node:
try:
gcolumns = h5file.createGroup('/', node_name, "Data %s" % node_name)
except tables.exceptions.NodeError:
h5file.removeNode('/', node_name, 1)
gcolumns = h5file.createGroup('/', node_name, "Data %s" % node_name)
elif node_name in h5file.root:
return h5file, getattr(h5file.root, node_name)
else:
gcolumns = h5file.createGroup('/', node_name, "Data %s" % node_name)
if FaceBBoxDDMPytables.h5file is None:
FaceBBoxDDMPytables.h5file = h5file
start = 0 if start is None else start
stop = gcolumns.X.nrows if stop is None else stop
atom = tables.Float32Atom() if config.floatX == 'float32' else tables.Float64Atom()
filters = FaceBBoxDDMPytables.filters
x = h5file.createCArray(gcolumns, 'X', atom = atom, shape = ((stop - start, data.X.shape[1])),
title = "Images", filters = filters)
y = h5file.createTable(gcolumns, 'bboxes', BoundingBox,
title = "Face bounding boxes", filters = filters)
x[:] = data.X[start:stop]
bboxes = get_image_bboxes(slice(start, stop), data.bboxes)
y.append(bboxes)
if remove_old_node:
h5file.removeNode('/', "Data", 1)
h5file.renameNode('/', "Data", node_name)
h5file.flush()
return h5file, gcolumns
def apply(self,
dataset,
start=None,
stop=None,
is_face=True,
can_fit=False):
"""
is_face: Bool,
Flag that determines if we operate on face or nonface dataset.
"""
if is_face:
X = dataset.face_dataset.get_topological_view()
else:
X = dataset.nonface_dataset.get_topological_view()
num_topological_dimensions = len(X.shape) - 2
if num_topological_dimensions != len(self.patch_shape):
raise ValueError("ExtractGridPatches with "
+ str(len(self.patch_shape))
+ " topological dimensions called on"
+ " dataset with " +
str(num_topological_dimensions) + ".")
num_patches = X.shape[0]
max_strides = [X.shape[0] - 1]
for i in xrange(num_topological_dimensions):
patch_width = self.patch_shape[i]
data_width = X.shape[i + 1]
last_valid_coord = data_width - patch_width
if last_valid_coord < 0:
raise ValueError('On topological dimension ' + str(i) +
', the data has width ' + str(data_width) +
' but the requested patch width is ' +
str(patch_width))
stride = self.patch_stride[i]
if stride == 0:
max_stride_this_axis = 0
else:
max_stride_this_axis = last_valid_coord / stride
num_strides_this_axis = max_stride_this_axis + 1
max_strides.append(max_stride_this_axis)
num_patches *= num_strides_this_axis
# batch size
output_shape = [num_patches]
# topological dimensions
for dim in self.patch_shape:
output_shape.append(dim)
# number of channels
output_shape.append(X.shape[-1])
output = numpy.zeros(output_shape, dtype=X.dtype)
channel_slice = slice(0, X.shape[-1])
coords = [0] * (num_topological_dimensions + 1)
keep_going = True
i = 0
bboxes = dataset.face_dataset.bboxes
while keep_going:
args = [coords[0]]
for j in xrange(num_topological_dimensions):
coord = coords[j + 1] * self.patch_stride[j]
args.append(slice(coord, coord + self.patch_shape[j]))
args.append(channel_slice)
patch = X[args]
output[i, :] = patch
i += 1
# increment coordinates
j = 0
keep_going = False
while not keep_going:
if coords[-(j + 1)] < max_strides[-(j + 1)]:
coords[-(j + 1)] += 1
keep_going = True
else:
coords[-(j + 1)] = 0
if j == num_topological_dimensions:
break
j = j + 1
portion = slice(start, stop)
repeat_times = numpy.ones(X.shape[0], 1)
if is_face:
bbox_targets = get_image_bboxes(portion, bboxes)
outputmaps = convert_bboxes_guided(bbox_targets, (X.shape[0], X.shape[1]),
self.patch_shape,
area_ratio=self.area_ratio,
stride=stride)
else:
targets = numpy.zeros(output_shape[-1])
outputmaps = repeat_times * targets
patch_loc = numpy.arange(output_shape[-1])
patch_locs = repeat_times * patch_loc
if start is None and stop is None:
return output, outputmaps.flatten(), patch_locs.flatten()
elif start is None:
raise ValueError("You should give a start value.")
elif start is not None and stop is None:
return output[start], outputmaps.flatten(), patch_locs[start].flatten()
else:
return output[start:stop], outputmaps.flatten(), patch_locs[start:stop].flatten()
def apply(self,
dataset,
start=None,
stop=None,
is_face=True,
can_fit=False):
"""
is_face: Bool,
Flag that determines if we operate on face or nonface dataset.
"""
if is_face:
X = dataset.face_dataset.get_topological_view()
else:
X = dataset.nonface_dataset.get_topological_view()
num_topological_dimensions = len(X.shape) - 2
if num_topological_dimensions != len(self.patch_shape):
raise ValueError("ExtractGridPatches with " +
str(len(self.patch_shape)) +
" topological dimensions called on" +
" dataset with " +
str(num_topological_dimensions) + ".")
num_patches = X.shape[0]
max_strides = [X.shape[0] - 1]
for i in xrange(num_topological_dimensions):
patch_width = self.patch_shape[i]
data_width = X.shape[i + 1]
last_valid_coord = data_width - patch_width
if last_valid_coord < 0:
raise ValueError('On topological dimension ' + str(i) +
', the data has width ' + str(data_width) +
' but the requested patch width is ' +
str(patch_width))
stride = self.patch_stride[i]
if stride == 0:
max_stride_this_axis = 0
else:
max_stride_this_axis = last_valid_coord / stride
num_strides_this_axis = max_stride_this_axis + 1
max_strides.append(max_stride_this_axis)
num_patches *= num_strides_this_axis
# batch size
output_shape = [num_patches]
# topological dimensions
for dim in self.patch_shape:
output_shape.append(dim)
# number of channels
output_shape.append(X.shape[-1])
output = numpy.zeros(output_shape, dtype=X.dtype)
channel_slice = slice(0, X.shape[-1])
coords = [0] * (num_topological_dimensions + 1)
keep_going = True
i = 0
bboxes = dataset.face_dataset.bboxes
while keep_going:
args = [coords[0]]
for j in xrange(num_topological_dimensions):
coord = coords[j + 1] * self.patch_stride[j]
args.append(slice(coord, coord + self.patch_shape[j]))
args.append(channel_slice)
patch = X[args]
output[i, :] = patch
i += 1
# increment coordinates
j = 0
keep_going = False
while not keep_going:
if coords[-(j + 1)] < max_strides[-(j + 1)]:
coords[-(j + 1)] += 1
keep_going = True
else:
coords[-(j + 1)] = 0
if j == num_topological_dimensions:
break
j = j + 1
portion = slice(start, stop)
repeat_times = numpy.ones(X.shape[0], 1)
if is_face:
bbox_targets = get_image_bboxes(portion, bboxes)
outputmaps = convert_bboxes_guided(bbox_targets,
(X.shape[0], X.shape[1]),
self.patch_shape,
area_ratio=self.area_ratio,
stride=stride)
else:
targets = numpy.zeros(output_shape[-1])
outputmaps = repeat_times * targets
patch_loc = numpy.arange(output_shape[-1])
patch_locs = repeat_times * patch_loc
if start is None and stop is None:
return output, outputmaps.flatten(), patch_locs.flatten()
elif start is None:
raise ValueError("You should give a start value.")
elif start is not None and stop is None:
return output[start], outputmaps.flatten(
), patch_locs[start].flatten()
else:
return output[start:stop], outputmaps.flatten(
), patch_locs[start:stop].flatten()