def __init__(self, sigma, sigma_k, sigma_b, lamb, sigma_trans, sl, sr,
rotation_angle, canopy):
super(UniversalTransformer, self).__init__()
self.sigma = sigma
self.sigma_k = sigma_k
self.sigma_b = sigma_b
self.scaler = transforms.BrightnessScale(sigma_k)
self.brighter = transforms.BrightnessShift(sigma_b)
self.gaussian_adder = transforms.ExpGaussian(lamb)
self.translation_adder = transforms.Translational(canopy, sigma_trans)
self.resize_adder = transforms.Resize(canopy, sl, sr)
self.rotation_adder = transforms.Rotation(canopy, rotation_angle)
self.noise_adder = transforms.Noise(self.sigma)
self.round = 1
self.masking = True
datasets = ['cifar10', 'mnist']
nums = [0, 20, 40, 60, 80, 100]
if __name__ == '__main__':
for dataset in datasets:
ds = get_dataset(dataset, 'test')
canopy = ds[0][0]
# init transformers
noiseT = transforms.Noise(sigma=0.5)
rotationT = transforms.Rotation(canopy, rotation_angle=180.0)
translationT = transforms.Translational(canopy, sigma=5.0)
blackTranslationT = transforms.BlackTranslational(canopy, sigma=5.0)
brightnessShiftT = transforms.BrightnessShift(sigma=0.1)
brightnessScaleT = transforms.BrightnessScale(sigma=0.1)
sizeScaleT = transforms.Resize(canopy, sl=0.5, sr=5.0)
gaussianT = transforms.Gaussian(sigma=5.0)
for num in nums:
print(dataset, num)
transforms.visualize(ds[num][0], f'visualize/{dataset}/{num}.png')
# rotation
angles = [-10, 30, 70]
for angle in angles:
transforms.visualize(
rotationT.masking(rotationT.raw_proc(ds[num][0], angle)),
f'visualize/{dataset}/{num}_rot_{angle}.png')
def gen_transform_and_params(args, canopy):
# init transformers
tinst1 = None
tfunc1 = None
tinst2 = None
tfunc2 = None
if args.transtype == 'gaussian':
tinst1 = TR.Gaussian(0.0)
tfunc1 = TR.Gaussian.proc
elif args.transtype == 'translation':
tinst1 = TR.Translational(canopy, 0.0)
tfunc1 = TR.Translational.proc
elif args.transtype == 'brightness':
tinst1 = TR.BrightnessShift(0.0)
tfunc1 = TR.BrightnessShift.proc
elif args.transtype == 'brightness-contrast':
# note: contrast is in exponential scale
tinst1 = TR.BrightnessShift(0.0)
tfunc1 = TR.BrightnessShift.proc
tinst2 = TR.BrightnessScale(0.0)
tfunc2 = TR.BrightnessScale.proc
elif args.transtype == 'rotation':
tinst1 = TR.Rotation(canopy, 0.0)
tfunc1 = TR.Rotation.raw_proc
elif args.transtype == 'scaling':
# note: resize is in original scale
tinst1 = TR.Resize(canopy, 1.0, 1.0)
tfunc1 = TR.Resize.proc
elif args.transtype == 'rotation-brightness' or args.transtype == 'rotation-brightness-l2':
tinst1 = TR.Rotation(canopy, 0.0)
tfunc1 = TR.Rotation.raw_proc
tinst2 = TR.BrightnessShift(0.0)
tfunc2 = TR.BrightnessShift.proc
elif args.transtype == 'scaling-brightness' or args.transtype == 'scaling-brightness-l2':
# note: resize is in original scale
tinst1 = TR.Resize(canopy, 1.0, 1.0)
tfunc1 = TR.Resize.proc
tinst2 = TR.BrightnessShift(0.0)
tfunc2 = TR.BrightnessShift.proc
# random generator
param1l, param1r, param2l, param2r, candidates = None, None, None, None, None
if args.transtype == 'gaussian':
param1l, param1r = 0.0, args.blur_alpha
elif args.transtype == 'translation':
candidates = torch.tensor(
list(
set([(x, y) for x in range(int(args.displacement) + 1)
for y in range(int(args.displacement) + 1)
if float(x * x + y * y) <= args.displacement *
args.displacement] +
[(x, -y) for x in range(int(args.displacement) + 1)
for y in range(int(args.displacement) + 1)
if float(x * x + y * y) <= args.displacement *
args.displacement] +
[(-x, y) for x in range(int(args.displacement) + 1)
for y in range(int(args.displacement) + 1)
if float(x * x + y * y) <= args.displacement *
args.displacement] +
[(-x, -y) for x in range(int(args.displacement) + 1)
for y in range(int(args.displacement) + 1)
if float(x * x + y * y) <= args.displacement *
args.displacement])))
c_len = candidates.shape[0]
param1l, param1r = 0.0, c_len
elif args.transtype == 'brightness':
param1l, param1r = -args.b, +args.b
elif args.transtype == 'brightness-contrast':
param1l, param1r = -args.b, +args.b
param2l, param2r = math.log(1.0 - args.k), math.log(1.0 + args.k)
elif args.transtype == 'rotation':
param1l, param1r = -args.r, +args.r
elif args.transtype == 'scaling':
param1l, param1r = 1.0 - args.s, 1.0 + args.s
elif args.transtype == 'rotation-brightness' or args.transtype == 'rotation-brightness-l2':
param1l, param1r = -args.r, +args.r
param2l, param2r = -args.b, +args.b
elif args.transtype == 'scaling-brightness' or args.transtype == 'scaling-brightness-l2':
param1l, param1r = 1.0 - args.s, 1.0 + args.s
param2l, param2r = -args.b, +args.b
else:
raise Exception(f'Unknown transtype: {args.transtype}')
print(f"""
param1: [{param1l}, {param1r}]
param2: [{param2l}, {param2r}]
""")
return tinst1, tfunc1, tinst2, tfunc2, param1l, param1r, param2l, param2r, candidates
# iterate through the dataset
dataset = get_dataset(args.dataset, args.split)
# init transformers
tinst = None
tfunc = None
tinst2 = None
tfunc2 = None
tinst3 = None
tfunc3 = None
if args.transtype == 'gaussian':
tinst = T.Gaussian(0.0)
tfunc = T.Gaussian.proc
elif args.transtype == 'translation':
tinst = T.Translational(dataset[0][0], 0.0)
tfunc = T.Translational.proc
elif args.transtype == 'brightness':
tinst = T.BrightnessShift(0.0)
tfunc = T.BrightnessShift.proc
elif args.transtype == 'contrast':
# note: contrast is in exponential scale
tinst = T.BrightnessScale(0.0)
tfunc = T.BrightnessScale.proc
elif args.transtype == 'brightness-contrast':
tinst = T.BrightnessShift(0.0)
tfunc = T.BrightnessShift.proc
tinst2 = T.BrightnessScale(0.0)
tfunc2 = T.BrightnessScale.proc
elif args.transtype == 'rotation':
tinst = T.Rotation(dataset[0][0], 0.0)
def __init__(self, sigma, canopy):
super(TranslationTransformer, self).__init__()
self.translation_adder = transforms.Translational(canopy, sigma)
self.sigma = sigma