def __init__(self, axes):
"""
Mirror data across specified axes, randomly during pre-processing, and using all permutations during full
pre-processing.
:param axes: Dimensions across which data is mirrored. (list)
"""
# Check first dimension (batch size) is not flipped
if 0 in axes:
raise ValueError('Cannot flip across dimension 0 (batch size)!')
# Prepare amount of crops and size at each dimension
self.axes = axes
# Call parent initializer
Operator.__init__(self, num_modes=2**len(self.axes))
def __init__(self, sizes, amounts=None):
"""
Crop data across all dimensions, randomly during random pre-processing, and at N evenly spread locations per
dimension during full pre-processing.
:param sizes: Size of the crop at each dimension. Use -1 to specify full size. (list)
:param amounts: Number of evenly-spread crops per dimension during full pre-processing. (list)
"""
# Prepare amount of crops and size at each dimension
amounts = [1] * len(sizes) if amounts is None else amounts
self.amounts = amounts if isinstance(
amounts, (list, tuple)) else [amounts] * len(self.sizes)
self.sizes = sizes
# Call parent initializer
Operator.__init__(self, num_modes=np.prod(amounts))
def __init__(self, axes, ranges):
"""
Shift data across specified axes, randomly during pre-processing, doing nothing during full-augmentation.
:param axes: Dimensions across which to shift. (list)
:param ranges: Range of the shifts across each dimension. (list of lists)
"""
# Check first dimension (batch size) is not shifted
if 0 in axes:
raise ValueError('Cannot shift across dimension 0 (batch size)!')
# Prepare operator attributes
self.axes = axes
self.ranges = ranges
# Call parent initializer
Operator.__init__(self, num_modes=1)
def __init__(self, order):
self.order = tuple(order)
Operator.__init__(self, num_modes=1)
def __init__(self, args, sess):
Operator.__init__(self, args, sess)
def __init__(self, weight=1, bias=0):
self.weight = weight
self.bias = bias
Operator.__init__(self, num_modes=1)
