Reconfigurable Intelligent Surfaces (RIS) represent an advanced technology reshaping wireless communication networks. Through intelligent configuration of wireless propagation environments using both cost-effective passive and active elements, RIS has the potential to bring about considerable performance improvements. In RIS-MIMO systems, precise control of passive RIS elements is crucial in optimizing reflected signal phases. This control necessitates intricate algorithms, given that inaccurate phase optimization can result in suboptimal signal focus and decreased data transmission accuracy. Obtaining accurate channel state information (CSI) is vital for achieving optimal phase control and high data rates; however, estimating channels between the transmitter, RIS, and receiver poses challenges. This paper investigates deep learning methodologies for channel estimation, explicitly addressing the distinctive challenges associated with phase shifts and beamforming. We present tailored deep-learning algorithms for each estimation technique, showcasing notable improvements in estimation accuracy, computational efficiency, and adaptability in dynamic environments.