The task of estimating optical flow is one of the unsolved problems in the field of computer vision. Optical flow has a wide application in practice and therefore there are numerous attempts to solve this problem. Starting from the formulation of the optical flow as an optimization problem and the use of hand-crafted descriptors, the current state-of-the-art in the field of optical flow has been achieved by deep convolutional models. An example of such a deep convolutional model is RAFT which consists of three main components: a feature encoder, a correlation layer and an update operator. The main feature of RAFT is the iterative approach and the use of cost volume with all pairs of pixels. In this paper, we propose a method for self-supervised learning of the RAFT model. We perform the learning in two phases by using the flow estimation in the non-occluded pixels from the first phase as ground-truth to guide learning in the manually occluded areas during the second phase. The proposed method is able to learn on pairs of images without knowing the ground-truth flow, thus removing restrictions on the field of application. We evaluate the estimations at the Sintel and Flying Chairs and show that our model achieves results comparable to the best known models, and we also discuss the possibilities for the future work.