def apply_policy(policy,
data,
image_size,
style_augmentor=None,
verbose=False):
"""
Apply the `policy` to the numpy `img`.
Args:
policy: A list of tuples with the form (name, probability, level) where
`name` is the name of the augmentation operation to apply, `probability`
is the probability of applying the operation and `level` is what strength
the operation to apply.
data: Numpy image list that will have `policy` applied to it.
image_size: (height, width) of image.
count: number of operations to apply
style_augmentor: function that augments data with randomized style
verbose: Whether to print applied augmentations.
Returns:
The result of applying `policy` to `data`.
"""
# PBA cifar10 policy modified
# count = np.random.choice([0, 1, 2, 3], p=[0.10, 0.20, 0.30, 0.40]) # old
count = np.random.choice([0, 1, 2, 3], p=[0.20, 0.20, 0.50, 0.10]) # new
#count = np.random.choice([0, 1, 2], p=[0.30, 0.50, 0.20]) # new2
# count = np.random.choice([0, 1, 2, 3], p=[0.2, 0.3, 0.5, 0.0]) # original
if count != 0:
data['img_data'] = pil_wrap(data['img_data'])
policy = copy.copy(policy)
random.shuffle(policy)
for xform in policy:
assert len(xform) == 3
name, probability, level = xform
assert 0. <= probability <= 1.
assert 0 <= level <= PARAMETER_MAX
xform_fn = NAME_TO_TRANSFORM[name].pil_transformer(
probability, level, image_size, style_augmentor)
data, res = xform_fn(data)
if verbose and res:
tf.logging.info("Op: {}, Magnitude: {}, Prob: {}".format(
name, level, probability))
count -= res
assert count >= 0
if count == 0:
break
data['img_data'] = pil_unwrap(
data['img_data'], image_size) # apply pil_unwrap on imgs only
return data
else:
return data
def apply_policy(policy, img, aug_policy, dset, image_size, verbose=False):
"""Apply the `policy` to the numpy `img`.
Args:
policy: A list of tuples with the form (name, probability, level) where
`name` is the name of the augmentation operation to apply, `probability`
is the probability of applying the operation and `level` is what strength
the operation to apply.
img: Numpy image that will have `policy` applied to it.
aug_policy: Augmentation policy to use.
dset: Dataset, one of the keys of MEANS or STDS.
image_size: Width and height of image.
verbose: Whether to print applied augmentations.
Returns:
The result of applying `policy` to `img`.
"""
if aug_policy == 'cifar100':
count = np.random.choice([0, 1, 2, 3], p=[0.2, 0.3, 0.5, 0.0])
else:
raise ValueError('Unknown aug policy.')
if count != 0:
pil_img = pil_wrap(img, dset)
policy = copy.copy(policy)
random.shuffle(policy)
for xform in policy:
assert len(xform) == 3
name, probability, level = xform
assert 0. <= probability <= 1.
assert 0 <= level <= PARAMETER_MAX
xform_fn = NAME_TO_TRANSFORM[name].pil_transformer(
probability, level, image_size)
pil_img, res = xform_fn(pil_img)
if verbose and res:
print("Op: {}, Magnitude: {}, Prob: {}".format(
name, level, probability))
count -= res
assert count >= 0
if count == 0:
break
return pil_unwrap(pil_img, dset, image_size)
else:
return img
