def do_augmentation(array):
"""Augmentation for the training data.
:array: A numpy array of size [c, x, y, z]
:returns: augmented image and the corresponding mask
"""
# normalize image to range [0, 1], then apply this transform
patch_size = np.asarray(array.shape)[1:]
augmented = noise_transforms.augment_gaussian_noise(array,
noise_variance=(0,
.015))
# need to become [bs, c, x, y, z] before augment_spatial
augmented = augmented[None, ...]
# mask = mask[None, None, ...]
r_range = (0, (3 / 360.) * 2 * np.pi)
cval = 0.
augmented, _ = spatial_transforms.augment_spatial(
augmented,
patch_size=patch_size,
seg=None,
do_rotation=True,
angle_x=r_range,
angle_y=r_range,
angle_z=r_range,
do_scale=True,
scale=(.9, 1.1),
border_mode_data='constant',
border_cval_data=cval,
order_data=3,
p_scale_per_sample=.5,
p_rot_per_sample=.5,
random_crop=False)
return augmented[0]
def do_augmentation(self, array, mask):
#array = array[None, ...]
patch_size = np.asarray(array.shape)
augmented = noise_transforms.augment_gaussian_noise(
array, noise_variance=(0, .015))
# need to become [bs, c, x, y, z] before augment_spatial
augmented = augmented[None, None, ...]
mask = mask[None, None, ...]
r_range = (0, (15 / 360.) * 2 * np.pi)
r_range2 = (0, (3 / 360.) * 2 * np.pi)
cval = 0.
augmented, mask = spatial_transforms.augment_spatial(
augmented,
seg=mask,
patch_size=patch_size,
do_elastic_deform=True,
alpha=(0., 100.),
sigma=(8., 13.),
do_rotation=True,
angle_x=r_range2,
angle_y=r_range2,
angle_z=r_range,
do_scale=True,
scale=(.9, 1.1),
border_mode_data='constant',
border_cval_data=cval,
order_data=1,
p_el_per_sample=0.5,
p_scale_per_sample=.5,
p_rot_per_sample=.5,
random_crop=False)
mask = mask[0]
augmented = (augmented[0, 0, :, :, :]).astype(np.uint8)
return augmented, mask