The model itself is written in keras, but to prepare vector labels I used scipy.ndimage.morphology
scipy.ndimage.morphology.distance_transform_edt() is very useful to find distance and direction to the border as it can return the index of the nearest pixel outside of mask:
edt, inds = scipy.ndimage.morphology.distance_transform_edt(crop_smooth, return_distances=True, return_indices=True)
border_vector = np.array([
np.expand_dims(np.arange(0, rows), axis=1) - inds[0],
np.expand_dims(np.arange(0, cols), axis=0) - inds[1]])
border_vector_norm = border_vector / (np.linalg.norm(border_vector, axis=0, keepdims=True) + 1e-5)
res_crop[:, :, 0] = border_vector_norm[0]
res_crop[:, :, 1] = border_vector_norm[1]
Vector to the center is easier to calculate, it's a difference of the pixel position and the mask center of mass:
center_of_mass = scipy.ndimage.measurements.center_of_mass(crop)
current_offset_field = np.zeros((CROP_SIZE, CROP_SIZE, 2))
current_offset_field[:, :, 0] = np.expand_dims(center_of_mass[0] - np.arange(0, CROP_SIZE), axis=1)
current_offset_field[:, :, 1] = np.expand_dims(center_of_mass[1] - np.arange(0, CROP_SIZE), axis=0)