def transform_source_batch(self, source, source_name):
self.verify_axis_labels(('batch', 'channel', 'height', 'width'),
self.data_stream.axis_labels[source_name],
source_name)
windowed_height, windowed_width = self.window_shape
if (isinstance(source, list) or
(isinstance(source, numpy.ndarray) and source.ndim == 1)) and all(
isinstance(b, numpy.ndarray) and b.ndim == 3 for b in source):
examples = [
self.transform_source_example(im, source_name) for im in source
]
if isinstance(source, list):
return examples
else:
return numpy.array(examples)
elif isinstance(source, numpy.ndarray) and source.ndim == 4:
# Hardcoded assumption of (batch, channels, height, width).
# This is what the fast Cython code supports.
batch_size = source.shape[0]
image_height, image_width = source.shape[2:]
if self.center_crop: # deterministic center crop
offset_y = (image_height - windowed_height) // 2
offset_x = (image_width - windowed_width) // 2
out = source[:, :, offset_y:-offset_y, offset_x:-offset_x]
else: # random crop
out = numpy.empty(source.shape[:2] + self.window_shape,
dtype=source.dtype)
max_h_off = image_height - windowed_height
max_w_off = image_width - windowed_width
if max_h_off < 0 or max_w_off < 0:
raise ValueError(
"Got ndarray batch with image dimensions {} but "
"requested window shape of {}".format(
source.shape[2:], self.window_shape))
offsets_w = self.rng.random_integers(0,
max_w_off,
size=batch_size)
offsets_h = self.rng.random_integers(0,
max_h_off,
size=batch_size)
window_batch_bchw(source, offsets_h, offsets_w, out)
if self.random_lr_flip:
for example in out:
if random.randint(0, 1):
example[:] = example[:, :, ::-1]
out = out.astype(numpy.float32)
if self.devide_by_255:
out = out.astype(numpy.float32) / 255.0
return out
else:
raise ValueError("uninterpretable batch format; expected a list "
"of arrays with ndim = 3, or an array with "
"ndim = 4")
def transform_source_batch(self, source, source_name):
self.verify_axis_labels(('batch', 'channel', 'height', 'width'),
self.data_stream.axis_labels[source_name],
source_name)
windowed_height, windowed_width = self.window_shape
if isinstance(source, numpy.ndarray) and source.ndim == 4:
# Hardcoded assumption of (batch, channels, height, width).
# This is what the fast Cython code supports.
out = numpy.empty(source.shape[:2] + self.window_shape,
dtype=source.dtype)
batch_size = source.shape[0]
# If padding is requested, pad before making random crop.
if self.pad is not None:
symmetric = (self.pad, self.pad)
paddims = ((0, 0), (0, 0), symmetric, symmetric)
source = numpy.pad(source, paddims, 'constant')
image_height, image_width = source.shape[2:]
max_h_off = image_height - windowed_height
max_w_off = image_width - windowed_width
if max_h_off < 0 or max_w_off < 0:
raise ValueError("Got ndarray batch with image dimensions {} "
"but requested window shape of {}".format(
source.shape[2:], self.window_shape))
offsets_w = self.rng.random_integers(0, max_w_off, size=batch_size)
offsets_h = self.rng.random_integers(0, max_h_off, size=batch_size)
window_batch_bchw(source, offsets_h, offsets_w, out)
# If flipping is requested, randomly flip images horizontally
if self.x_flip:
whichflip = self.rng.binomial(1, 0.5, batch_size)
out[whichflip, :, :, :] = out[whichflip, :, :, ::-1]
return out
elif all(isinstance(b, numpy.ndarray) and b.ndim == 3 for b in source):
return [
self.transform_source_example(im, source_name) for im in source
]
else:
raise ValueError("uninterpretable batch format; expected a list "
"of arrays with ndim = 3, or an array with "
"ndim = 4")
def transform_source_batch(self, source, source_name):
self.verify_axis_labels(('batch', 'channel', 'height', 'width'),
self.data_stream.axis_labels[source_name],
source_name)
windowed_height, windowed_width = self.window_shape
if isinstance(source, numpy.ndarray) and source.ndim == 4:
# Hardcoded assumption of (batch, channels, height, width).
# This is what the fast Cython code supports.
out = numpy.empty(source.shape[:2] + self.window_shape,
dtype=source.dtype)
batch_size = source.shape[0]
# If padding is requested, pad before making random crop.
if self.pad is not None:
symmetric = (self.pad, self.pad)
paddims = ((0, 0), (0, 0), symmetric, symmetric)
source = numpy.pad(source, paddims, 'constant')
image_height, image_width = source.shape[2:]
max_h_off = image_height - windowed_height
max_w_off = image_width - windowed_width
if max_h_off < 0 or max_w_off < 0:
raise ValueError("Got ndarray batch with image dimensions {} "
"but requested window shape of {}".format(
source.shape[2:], self.window_shape))
offsets_w = self.rng.random_integers(0, max_w_off, size=batch_size)
offsets_h = self.rng.random_integers(0, max_h_off, size=batch_size)
window_batch_bchw(source, offsets_h, offsets_w, out)
# If flipping is requested, randomly flip images horizontally
if self.x_flip:
whichflip = self.rng.binomial(1, 0.5, batch_size)
out[whichflip,:,:,:] = out[whichflip,:,:,::-1]
return out
elif all(isinstance(b, numpy.ndarray) and b.ndim == 3 for b in source):
return [self.transform_source_example(im, source_name)
for im in source]
else:
raise ValueError("uninterpretable batch format; expected a list "
"of arrays with ndim = 3, or an array with "
"ndim = 4")