def __init__(self,
img_shape,
blur_sigma,
flow_sigma,
normalize_max=False,
**kwargs):
super(RandFlow, self).__init__(**kwargs)
n_dims = len(img_shape) - 1
self.flow_shape = img_shape[:-1] + (n_dims, )
if blur_sigma > 0:
blur_kernel = image_utils.create_gaussian_kernel(blur_sigma,
n_dims=n_dims)
# TODO: make this work for 3D
if n_dims == 2:
blur_kernel = np.tile(
np.reshape(blur_kernel, blur_kernel.shape + (1, 1)),
tuple([1] * n_dims) + (n_dims, 1))
else:
blur_kernel = np.reshape(blur_kernel,
blur_kernel.shape + (1, 1))
self.blur_kernel = tf.constant(blur_kernel, dtype=tf.float32)
else:
self.blur_kernel = None
self.flow_sigma = flow_sigma
self.normalize_max = normalize_max
self.n_dims = n_dims
print('Randflow dims: {}'.format(self.n_dims))
def __init__(self, img_shape, blur_sigma, **kwargs):
super(BlurFlow, self).__init__(**kwargs)
n_dims = len(img_shape) - 1
self.flow_shape = tuple(img_shape[:-1]) + (n_dims,)
blur_kernel = image_utils.create_gaussian_kernel(blur_sigma, n_dims=n_dims, n_sigmas_per_side=2)
# TODO: make this work for 3D
if n_dims == 2:
blur_kernel = np.tile(np.reshape(blur_kernel, blur_kernel.shape + (1, 1)),
tuple([1] * n_dims) + (n_dims, 1))
else:
blur_kernel = np.reshape(blur_kernel, blur_kernel.shape + (1, 1))
self.blur_kernel = tf.constant(blur_kernel, dtype=tf.float32)
self.n_dims = n_dims
def __init__(self, img_shape, dilate_kernel_size, blur_sigma, flow_sigma, **kwargs):
super(DilateAndBlur, self).__init__(**kwargs)
n_dims = len(img_shape) - 1
self.flow_shape = img_shape[:-1] + (n_dims,)
dilate_kernel = np.reshape(
cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (dilate_kernel_size, dilate_kernel_size)),
(dilate_kernel_size, dilate_kernel_size, 1, 1))
self.dilate_kernel = tf.constant(dilate_kernel, dtype=tf.float32)
blur_kernel = image_utils.create_gaussian_kernel(blur_sigma, n_dims=n_dims)
blur_kernel = np.reshape(blur_kernel, blur_kernel.shape + (1, 1))
blur_kernel = blur_kernel / np.max(blur_kernel) # normalize by max instead of by sum
self.blur_kernel = tf.constant(blur_kernel, dtype=tf.float32)
self.flow_sigma = flow_sigma
def __init__(self, n_chans=3, n_dims=2, do_blur=True, **kwargs):
super(Blur_Downsample, self).__init__(**kwargs)
scale_factor = 0.5 # we only support halving right now
if do_blur:
# according to scikit-image.transform.rescale documentation
blur_sigma = (1. - scale_factor) / 2
blur_kernel = image_utils.create_gaussian_kernel(blur_sigma, n_dims=n_dims, n_sigmas_per_side=4)
# if n_dims==2:
# blur_kernel = np.tile(np.reshape(blur_kernel, blur_kernel.shape + (1, 1)), tuple([1] * n_dims) + (n_chans, 1))
# else:
blur_kernel = np.reshape(blur_kernel, blur_kernel.shape + (1, 1))
self.blur_kernel = tf.constant(blur_kernel, dtype=tf.float32)
else:
self.blur_kernel = tf.constant(np.ones([1] * n_dims + [1, 1]), dtype=tf.float32)
self.n_dims = n_dims
self.n_chans = n_chans
def __init__(self, n_chans, n_dims, do_blur=True, **kwargs):
super(Blur_Downsample, self).__init__(**kwargs)
scale_factor = 0.5 # we only support halving right now
if do_blur:
# according to scikit-image.transform.rescale documentation
blur_sigma = (1. - scale_factor) / 2
blur_kernel = image_utils.create_gaussian_kernel(blur_sigma, n_dims=n_dims, n_sigmas_per_side=4)
if n_dims==2:
blur_kernel = np.tile(np.reshape(blur_kernel, blur_kernel.shape + (1,1)), tuple([1]*n_dims) + (n_dims, 1))
else:
blur_kernel = np.reshape(blur_kernel, blur_kernel.shape + (1,1))
self.blur_kernel = tf.constant(blur_kernel, dtype=tf.float32)
else:
self.blur_kernel = tf.constant(np.ones([1] * n_dims + [1, 1]), dtype=tf.float32)
self.n_dims = n_dims
self.n_chans = n_chans
