def get_instance(rnd):
U = rnd.uniform
N = rnd.normal
rr = rotation_range
hs = height_shift_range
ws = width_shift_range
sr = shear_range
agn = additive_gaussian_noise_range
dp = distortion_prob
mg = multiplication_gaussian
tcp = transform_colorspace_param
tcb = transform_colorspace_bounds
return {
'input_shape': input_shape,
'theta': np.pi / 180 * U(-rr, rr) if rr else 0,
'ty': U(-hs, hs) * input_shape[0] if hs else 0,
'tx': U(-ws, ws) * input_shape[1] if ws else 0,
'shear': U(-sr, sr) if shear_range else 0,
'z': U(zoom_range[0], zoom_range[1]) if zoom_range != (1, 1) else 1,
'h_flip': rnd.rand() < 0.5 if horizontal_flip else False,
'add_noise': N(0, U(agn[0], agn[1]), input_shape) if agn is not None else None,
'resize': U(*resize_range) if U(0, 1) < dp else None,
'resize_smooth': U(0, 1) < 0.5,
'mul': N(1, mg) if mg > 0 else None,
'color_m': utils.crop_value(N(tcp[0], tcp[1], (3, 3)), tcb) if tcp is not None else None,
'agn': agn
}
def transform(v,
t_matrix,
h_flip=False,
add_noise=None,
resize=None,
resize_smooth=None,
mul=None,
color_m=None):
"""
Transform image with (inverted) transformation matrix
:param v: Input image to be transformed
:param t_matrix: Transformation matrix
:param h_flip: Whether do horizontal flip
:return: Transformed image
"""
def apply_transform(x,
transform_matrix,
channel_index=0,
fill_mode='nearest',
cval=0.):
x = np.rollaxis(x, channel_index, 0)
final_affine_matrix = transform_matrix[:2, :2]
final_offset = transform_matrix[:2, 2]
channel_images = [
ndimage.interpolation.affine_transform(x_channel,
final_affine_matrix,
final_offset,
order=2,
mode=fill_mode,
cval=cval)
for x_channel in x
]
x = np.stack(channel_images, axis=0)
x = np.rollaxis(x, 0, channel_index + 1)
return x
def flip_axis(x, axis):
x = np.asarray(x).swapaxes(axis, 0)
x = x[::-1, ...]
x = x.swapaxes(0, axis)
return x
v = apply_transform(v, t_matrix, 2)
if h_flip:
v = flip_axis(v, 1)
if color_m is not None or mul is not None or add_noise is not None:
v = v.astype(np.float32)
shape = v.shape
if mul is not None:
v *= mul
if color_m is not None:
v = np.reshape(v, [-1, 3])
v = np.matmul(v, color_m)
v = np.reshape(v, shape)
if add_noise is not None:
v += add_noise
if resize is not None:
interpolation = 'bilinear' if resize_smooth else 'nearest'
v = imresize(v, (resize * np.array(shape[:2])).astype(np.uint16),
interpolation)
v = imresize(v, shape[:2], interpolation)
v = utils.crop_value(v, [np.zeros(shape), np.ones(shape) * 255])
return v.astype(np.uint8)
if color_m is not None:
v = np.reshape(v, [-1, 3])
v = np.matmul(v, color_m)
v = np.reshape(v, shape)
if add_noise is not None:
v += add_noise
if resize is not None:
interpolation = 'bilinear' if resize_smooth else 'nearest'
v = imresize(v, (np.resize * np.array(shape[:2])).astype(np.uint16), interpolation)
v = imresize(v, shape[:2], interpolation)
v = utils.crop_value(v, [np.zeros(shape), np.ones(shape) * 255])
return v.astype(np.uint8)
def crop_data(img, labels, new_img_size, new_label_size = None, crop_label = True):
"""
Both images and labels will be cropped to match the given size
:param img: Images to be cropped
:param labels: Labels to be cropped
:param new_img_size: New image size
:param new_label_size: New labels size
:return: Cropped image and labels
"""